Ask us anything about AI agents

BobWork builds, runs and keeps updating an AI agent that answers your customers' messages on WhatsApp, Instagram, Messenger and web chat.

The important word is "runs". It is a managed service, not software you are handed. You do not configure it, maintain it or learn a dashboard — you tell us what changed and we make the change.

With software, you buy a tool and the work of making it good is yours: writing the knowledge, tuning the answers, fixing it when it gets things wrong, keeping it current when your prices change. Most chatbot projects die there — not because the tool was bad, but because nobody had time to run it.

With BobWork the work is ours. You get the outcome, not the toolkit. That is why there is a setup fee and a monthly retainer rather than a per-seat licence.

No. Button flows break the moment a customer types something the flow did not anticipate, which is most of the time.

The agent reads the message, works out what the person actually wants, finds the answer in your business knowledge, and writes a reply in your own words. If it cannot answer confidently, it hands the conversation to a person instead of guessing.

Businesses getting more inbound messages than their team can answer well — typically clinics, e-commerce stores, restaurants, tuition centres, property agencies, travel companies and professional services firms.

The common symptom is not volume alone. It is that replies are slow after hours, the same twenty questions are answered by hand every day, and enquiries are lost because nobody got to them in time.

Two to four weeks for most deployments. The variable is almost never our build time — it is how quickly your business knowledge (prices, policies, product details) can be gathered, and how long Meta takes to verify your WhatsApp Business account.

If you already have a verified WhatsApp Business number and your information is written down somewhere, the fast end is realistic.

Three things: whatever you already have written down about your business (price lists, FAQs, policies, product data, past chat exports), access to the channel you want it on, and one person on your side who can answer questions and make decisions.

You do not need a tidy knowledge base. Assembling and structuring that is our job — messy source material is normal and expected.

Yes. Ask for a demo and we will show an agent handling real conversations end to end, then walk through what one built on your business would look like.

See bobwork.ai/demo, or go straight to bobwork.ai/talk-to-sales and tell us your channels and volume.

Very little. The agent sits on the channels you already use and answers the messages your team would otherwise answer by hand.

What changes is what reaches your team: the routine questions stop arriving, and what does arrive comes with the conversation history attached. Most teams change nothing except how many chats they open.

For each message it works out what the customer is asking, searches your business knowledge for the passages that bear on that question, and writes an answer grounded in what it found — in your tone, in the customer's language.

It is not picking from a list of pre-written replies, and it is not free-associating either. The knowledge you give it is the boundary.

We use commercial large language models from established providers, and we change which one as the field moves — a model that was best six months ago usually is not today.

That choice is ours to manage, deliberately. You are buying an agent that works, not a particular model number, and we re-tune when we switch so the behaviour you signed off on stays the same.

It is built specifically not to. The agent answers from your knowledge base rather than from general world knowledge, and when it cannot find an answer there it is instructed to say so and escalate rather than fill the gap.

The honest version: no AI system is perfect, which is why the escalation rules, the review period after launch and the "stop ai" controls exist. In practice the wrong answers we see come from knowledge that is missing or out of date, not from invention — which is why we ask you to tell us the moment something changes.

It says so plainly and hands the conversation to a person, rather than guessing or stalling.

We would rather it escalate too often in the first weeks than too little. Every escalation is also a signal: it tells us exactly which piece of knowledge is missing, and we add it.

Not silently, no — and that is deliberate. An agent that quietly rewrites its own behaviour based on whatever customers said this week is an agent nobody can predict.

Improvement happens through review: we look at what it got wrong, correct the underlying knowledge or tuning, and the change takes effect. You can see what changed and when.

Yes, within the conversation. It follows context across messages, so a customer can say "how much is that one?" after asking about a product and it knows what "that one" means.

It also has the history of that contact's earlier conversations available, so a returning customer does not start from zero.

Seconds, at any hour. That is the biggest single change most businesses notice — not that the answers are cleverer than a person's, but that they arrive at 11pm on a Sunday instead of at 10am on Monday.

It can act, through what we call skills: qualifying a lead, booking a meeting, generating a quotation, recommending products, checking an order's status, running follow-up sequences.

Your plan includes any three skills. Others can be added at any point without rebuilding the agent.

That is your decision, and we will implement whichever you choose. What the agent will not do is claim to be human when a customer asks directly — it is instructed to be straight about that and to offer a person.

Most of our customers find that disclosure costs them nothing. Speed and a correct answer is what people actually want.

Yes, without limit and without degrading. A hundred people messaging at the same moment each get an immediate reply. That is the point of the thing: your busiest hour stops being your worst hour.

Guardrails are part of setup. You tell us the subjects it must never handle — refunds beyond a threshold, medical or legal advice, negotiating discounts, anything commercially sensitive — and it escalates those to a person instead of attempting them.

This is one of the first things we agree, before it ever talks to a customer.

Yes, with four reply modes:

All chats — it answers every incoming message. Best for full 24/7 coverage.

New contacts only — it answers the first message from a brand-new contact; returning contacts go to your team.

Keyword triggers — it stays silent on a chat until the customer sends a keyword you choose. Once triggered, it stays active for that conversation.

Tagged chats — it only answers chats carrying tags you nominate. If you turn this on without adding any tags it replies to nobody rather than everybody, on purpose.

Tag that chat "stop ai". It is a built-in tag that works in every mode with no configuration, and the agent will not reply to that conversation.

It is the fastest way to pull the agent out of one chat — faster than changing the reply mode, which affects every conversation. You can also nominate your own excluded tags, which block replies in every mode too.

Tell us and we switch it off, or set the reply mode so it answers nothing. There is no lock-in mechanism that keeps it talking to your customers against your wishes.

Yes — every conversation is visible, and we expect you to read them, particularly in the first weeks.

That review is the most valuable thing you can do early on. It is how the gaps get found while the stakes are still low.

Yes, if that is what you want — you can restrict it to your opening hours so out-of-hours messages wait for your team.

Most customers do the opposite, because out-of-hours is where the unanswered messages actually are. Times run on Malaysia Time (GMT+8) by default and can be set to your timezone.

Tell us, with the conversation. We diagnose whether it was a knowledge gap (the right information was never given to it, or has gone stale) or a tuning problem (it had the information and handled it badly), and fix the cause rather than patching the symptom.

Fixes to knowledge take effect quickly. If the error had commercial consequences, say so — we will treat it as urgent rather than routine.

Commercially, it is your business speaking to your customer — the same as if a staff member had made the error, and the terms at bobwork.ai/terms set out how responsibility is shared.

Operationally, preventing it is our job: the guardrails, the escalation rules, the review period and the ongoing corrections all exist so this stays rare. If you are weighing a genuinely high-stakes use, tell us during scoping and we will scope the agent conservatively.

Only as much as you want it to. Tone and sales posture are set during setup and you sign off on them.

If you want it to answer helpfully and stop, it does that. If you want it to qualify and ask for the booking, it does that instead.

It is designed to recognise them and get out of the way. Frustration, complaints and anything emotionally charged are escalation triggers — the agent hands over to a person with the full conversation attached rather than trying to de-escalate on its own.

An AI arguing with an upset customer is a bad outcome for everyone, so it is built not to.

WhatsApp, Instagram, Messenger and web chat. They are all covered by the same plan and the same conversation allowance — there is no separate per-channel setup fee.

Yes, and it should. One agent, one set of knowledge, one tone across every channel — so a customer gets the same answer whether they message on Instagram or WhatsApp.

Maintaining separate bots per channel is how businesses end up contradicting themselves.

Yes — web chat is one of the included channels, embedded on your site as a widget.

It uses the same knowledge and the same tone as your messaging channels, so website visitors get identical answers to WhatsApp customers.

Instagram direct messages, yes — it is one of the four included channels, using the same knowledge as everything else.

If public comment handling matters to you, raise it during scoping so we can confirm what is possible against Meta's current rules rather than promise it here.

Not part of the standard four, so treat it as a scoping question rather than an assumption. Ask at bobwork.ai/talk-to-sales and we will tell you straight whether it is feasible, what it would take, and whether it is worth it.

We would rather say no clearly than yes vaguely.

Usually a dedicated number for the agent is cleanest, and we provision it as part of setup. Using your existing business number is often possible too.

Which is right depends on how your team currently uses that number, so we work it out during onboarding rather than prescribing it.

A WhatsApp Business account verified with Meta, a number provisioned for it, and — for messages you send first — templates approved by Meta.

We handle the process with you. The part outside anyone's control is Meta's own verification queue, which is usually the longest pole in the timeline.

Not for being AI. Meta's rules are about how you message people — sending unsolicited bulk messages, messaging people who never opted in, ignoring opt-outs. Those get numbers restricted whether a human or an agent sends them.

Answering customers who messaged you first, which is what the agent mostly does, is exactly what the platform is for. We set up outbound messaging to stay inside the rules, and we will tell you if something you want to do is likely to put your number at risk.

Meta's rule: once a customer messages you, you can reply freely for 24 hours. After that you can only reach them using a message template Meta has pre-approved.

It matters mainly for follow-ups. It is why follow-up sequences need approved templates and why timing them well is worth the thought.

Meta bills WhatsApp Business API conversations on its own schedule, separately from what you pay us, and the rates vary by country and conversation type.

We will walk you through what your volume is likely to cost at Meta's current rates during scoping, so there is no surprise line item later.

Yes — that is the normal arrangement. The agent handles what it should and your team steps into any conversation at any time, on the same number.

How that works in practice depends on your current WhatsApp setup, which we go through during onboarding.

A template is a pre-written message format Meta approves in advance, required for any message you send outside the 24-hour window.

Approval usually takes hours to a couple of days. We write and submit them for you, and we keep a set current so follow-ups are not blocked when you need them.

Four stages. We gather and structure your business knowledge. We connect the channels. We agree the behaviour — tone, guardrails, escalation rules, which conversations it handles. Then we test it against your real past conversations, not invented ones, and you sign off before it talks to anybody.

A few hours in total, spread over the build — mostly answering questions about how your business works and reviewing the agent's test answers before launch.

The review is the part worth doing properly. Everything else we can carry.

No, and it is the normal situation. Most of what a business knows lives in people's heads and in past conversations.

We work from whatever exists — a price list, a website, a WhatsApp history — and interview you for the rest. Exporting your past chats is often the single most useful thing you can hand over, because it shows what customers actually ask.

Yes, and it is usually the right way. Start on one channel with a clear scope, get it reliably right, then widen.

Skills can be added at any point without rebuilding the agent, so starting narrow costs you nothing later.

Deliberately unglamorous. We usually start it on a limited slice — new contacts only, or chats you tag — watch closely for a few days, correct what needs correcting, then widen the scope.

A quiet launch that you can inspect is worth more than a big-bang switch-on.

The retainer starts doing its work. We watch how it performs, correct what it gets wrong, and make every content or workflow change you ask for — live within 24 hours on weekdays.

This is the part that separates a managed agent from a tool. Businesses change constantly, and an agent that is not maintained goes stale within weeks.

Tell us what changed — new price, new policy, new product, a reply you want worded differently. Content and workflow changes go live within 24 hours on weekdays.

You do not need to describe it in technical terms or edit anything yourself.

Everything the agent is allowed to answer from: products, prices, policies, opening hours, delivery terms, common questions, and the answers your team already gives.

Its quality sets the ceiling on the agent's quality. A well-run agent with thin knowledge is still thin.

Anything a customer might ask and a new staff member would need told. Price lists, product details, policies on returns and cancellations, opening hours, locations, delivery times, the questions you answer twenty times a week.

The most valuable input is usually your past conversations: they show what customers actually ask, in their own words, rather than what you assume they ask.

5,000 knowledge files are included each month. Beyond that it is $10 per extra 1,000 files a month.

For most businesses the included allowance is not close to a constraint.

Tell us, and it is live within 24 hours on weekdays. You do not edit anything.

The thing worth building is the habit: whoever changes a price should have telling us on the same checklist as updating the website. Stale knowledge is the most common cause of wrong answers we see.

Yes — a website is a normal source, and often the starting point.

We usually do not stop there, because websites tend to be marketing-shaped rather than question-shaped. The gaps are the specific things customers ask that the site never quite says.

Yes, through skills and integrations, rather than from static knowledge. Order status and tracking is an available skill, and a Shopify connection lets it check real orders, stock and fulfilment inside the chat.

For other systems, tell us what you run and we will tell you honestly whether a live connection is feasible.

Good: specific, current, written the way a customer asks. "Standard delivery within Klang Valley is 2 working days; outside is 3 to 5."

Bad: vague, aspirational, or written for marketing. "We pride ourselves on fast delivery." The agent cannot turn the second into an answer, so it escalates a question you could have automated.

Four triggers: the customer is upset, the agent is not confident enough to answer, the request is commercial or sensitive in a way you have said needs a person, or the customer asks for a human.

The last one is absolute. If somebody asks for a person, they get one.

Your team is notified and picks up the conversation with the full history in front of them — so the customer never has to repeat themselves, which is the part that usually makes handovers feel bad.

Your team can also take over any conversation at any time without waiting for a trigger. The agent pauses on that chat until you resume it.

Not in our experience — it changes what they spend the day on. The repetitive questions stop reaching them; the conversations that need judgement still do, and arrive with context attached.

Most of our customers redeploy the time into work that was being crowded out, not into headcount cuts. If your intent is to reduce headcount, say so during scoping so we can be straight with you about what is realistic.

There is no per-seat charge, so this is not a licensing question. How many of your team work in the inbox is up to you.

Barely. They keep working in the same inbox they already use. The two things worth knowing are how to take over a conversation and how to tag a chat "stop ai" — both take a minute to explain.

Nobody on your side needs to learn to configure or maintain the agent, because nobody on your side does that.

Yes — routing to the right person or team is available as a skill, so enquiries can go to sales while support issues go elsewhere.

Tell us how your team is split during scoping and we will set the routing to match.

English, Malay and Chinese, including the mixed way people genuinely message in Malaysia and Singapore.

It detects the language and can switch mid-conversation, so a customer who opens in English and continues in Malay is followed rather than corrected.

Yes. Real messages are short, misspelled, mixed-language and full of local shorthand, and it is built for that rather than for tidy sentences.

We test against your actual past conversations during setup, which is where this gets proven rather than assumed.

Not if it is set up properly. Tone is configured deliberately — how formal, how long the replies run, whether it uses emoji, how it greets people.

We can also extract a style profile from your team's real conversations so it writes the way your people write, rather than the way a generic assistant writes.

From real conversations, not from a description. We take a chat where a team member replied the way you want the agent to sound, and extract a style profile from it — sentence length, punctuation, emoji use, how they greet people, how they handle a price question or a complaint.

You can hold several profiles, so a sales lead and a support lead can sound different, and refresh them whenever your team's style shifts.

Yes, at any time. It is a normal change request — tell us what you want different and it is live within 24 hours on weekdays.

Capabilities beyond answering questions. The library covers lead qualification, smart routing, quotation generation, product recommendations, meeting and booking, follow-up sequences, order status and tracking, and reviews and re-order prompts.

Your plan includes any three. Extras have a one-time setup fee — $500 small, $1,000 medium, $2,000 large — plus a small monthly fee, and can be added at any point.

Yes — meeting and booking is one of the skills. It picks up booking intent, collects what it needs, and confirms the appointment, including rescheduling and reminders.

For clinics this is usually the single highest-value thing to automate, because it removes a large share of reception's phone load.

Sometimes directly, sometimes not — it depends entirely on what you run and what it exposes.

Shopify is supported for orders, stock and fulfilment. For anything else, tell us the system during scoping and we will give you a straight answer rather than a hopeful one. Deeper system integration is also what the enterprise scope is for.

Treat it as a scoping question. What is straightforward is the step before payment — quoting, confirming what the customer wants, and sending them to your existing payment method.

Tell us how you take money today and we will tell you what is sensible to automate and what is not worth the risk.

Yes — follow-up sequences are a skill. Leads who did not reply get a follow-up automatically, at timing you agree.

On WhatsApp this is where Meta's 24-hour window applies, so follow-ups outside it use approved templates. We set that up rather than leaving you to discover it.

You do. Your business knowledge and your customer conversations are yours.

The formal terms are at bobwork.ai/terms and bobwork.ai/dpa. If your legal team needs specifics beyond what is written there, ask us and we will get you a direct answer rather than a marketing one.

Your business knowledge exists to answer your customers, not to improve anybody else's model.

This is a question that deserves precision rather than reassurance, and the exact terms — including what our providers may and may not do — are set out at bobwork.ai/privacy and bobwork.ai/dpa. If your compliance team needs it confirmed in writing for your contract, ask and we will put it in writing.

Conversations and business knowledge are held in access-controlled systems, and data moves over encrypted connections.

We would rather point you to the written commitments at bobwork.ai/privacy and bobwork.ai/dpa than make claims in a FAQ. If you have a security questionnaire, send it — a real answer to your actual questions is more useful than a paragraph here.

Ask us rather than assume, and we will tell you exactly where we stand rather than imply something we cannot back.

If a certification is a hard requirement for your procurement process, raise it early in the conversation. That is a better use of everyone's time than discovering it at contract stage.

Our data processing terms are at bobwork.ai/dpa, and that document is the right basis for your legal team's review.

For specific obligations in your jurisdiction or industry, put the question to us directly. Regulatory questions deserve a considered answer from a person, not a FAQ entry.

Ask us and we will tell you specifically for your deployment, including whether a particular region is possible if you have a residency requirement.

Worth raising early if data residency is a hard constraint for you, since it can affect how we set things up.

It is your data, and the terms at bobwork.ai/terms and bobwork.ai/dpa govern what happens at the end of an engagement, including deletion.

If you want specific commitments on export format or deletion timelines written into your contract, raise it during scoping and we will agree it up front.

A one-off setup fee to build the agent, then a monthly retainer to run and keep updating it.

Launch AI Agent — $1,500 setup, $1,000/month. The base plan: one managed AI agent, knowledge import, core replies with human handoff, 5,000 conversations a month, any 3 skills.

Sales Engine — $4,500 setup, $1,499/month. Adds lead capture and qualification, quoting and recommendations, automatic meeting booking, managed updates every weekday.

Full Funnel Engine — $8,500 setup, $1,999/month. Full sales automation from first message to reorder, including follow-up and reorder flows, order tracking and reviews.

Full detail at bobwork.ai/pricing.

Because the setup is real work, not an activation click. It covers importing and structuring your knowledge, connecting your channels, configuring the skills you picked, and testing the agent against your real conversations before it goes live.

That work is what makes the difference between an agent that holds up and one that embarrasses you in week two.

Running the agent: hosting, the included 5,000 conversations and 5,000 knowledge files, human handoff, and every content or workflow change you request — live within 24 hours on weekdays.

It is not a licence fee with support attached. It is the ongoing operation of the thing, which is the part that keeps it working as your business changes.

Your plan includes 5,000 a month. Above that, extra conversations are $0.05 each.

There is no per-message charge and no per-seat charge, so a long back-and-forth does not cost more than a short one.

Conversations above 5,000 a month are $0.05 each. Knowledge above 5,000 files is $10 per extra 1,000 files a month.

Overage is billed automatically via Stripe every 5 days, plus a final charge before your plan renews. There are no top-up packs to buy in advance — you pay for what you actually went over by.

Yes. Set an overage spending limit and AI replies pause once you reach it, rather than continuing to bill you.

If you run high or spiky volume, set it deliberately rather than leaving it open. It is the control that makes usage-based billing predictable.

Terms are agreed during scoping and set out at bobwork.ai/terms. Ask about notice periods and commitment length directly — you should have a clear answer before you sign, not after.

From us, no — setup, retainer, any extra skills, and usage above the allowance is the whole list, and all of it is on bobwork.ai/pricing.

The one cost that is not ours: Meta bills WhatsApp Business API conversations separately, at its own rates. We will walk you through what your volume implies before you commit, so it is not a surprise.

By card via Stripe. Connecting a card is also what enables overage to be billed — without it, usage above the allowance cannot be charged.

Yes. Custom builds start from $1,500 setup and $499/month, and the enterprise scope covers deeper system integration, multiple languages, compliance review cycles and dedicated support.

If your requirements do not fit the three packages, that is a conversation rather than a problem — bobwork.ai/talk-to-sales.

It depends far more on your business than on the AI. A clinic answering the same twenty questions automates most of them; a consultancy where every enquiry is bespoke automates far less.

We would rather scope your actual conversations and give you a grounded number than quote an industry average that means nothing for you. Ask during scoping.

Agree what you are measuring before launch, not after. Usually: how many conversations were handled without a person, how fast first replies went out, how many enquiries turned into bookings or orders, and how much of your team's day came back.

The most honest baseline is your own numbers from the month before launch.

The immediate, visible change is response time — messages answered in seconds at any hour, including the ones that used to sit overnight. That shows up in week one.

The commercial return depends on what those unanswered messages were worth to you. If you know roughly what an enquiry is worth and how many you were losing after hours, that arithmetic is the honest place to start.

Tell us early and specifically. Most "it is not working" turns out to be a scope or knowledge problem we can fix quickly, and the first weeks are exactly when it should be raised.

If it genuinely is not right for your business, we would rather establish that honestly than keep billing you. Commercial terms for ending an engagement are at bobwork.ai/terms.

Usually one of five things: the reply mode is narrower than you think (tagged chats mode with no tags configured replies to nobody, by design), the chat carries an excluded tag or the "stop ai" tag, somebody took the conversation over in Copilot mode and did not resume, business-hours-only is on and it is outside them, or a spending limit has been reached.

If none of those explain it, send us the conversation and we will look.

That is the reply mode. "New contacts only" answers the first message from a brand-new contact and leaves returning contacts to your team; "tagged chats" only answers chats carrying tags you nominated.

Tell us how you want it to behave and we will change the mode.

Send us the conversation and the correct price. Almost always the knowledge was never updated when the price changed.

The fix is live within 24 hours on weekdays, and worth pairing with a habit: whoever changes a price tells us at the same time they update the website.

That is a knowledge gap showing itself, and it is the failure mode you want — it escalated instead of inventing an answer.

Send us a few examples. Each one names a specific thing missing from the knowledge base, and adding them turns those escalations into automated answers.

Send us the conversation. Language detection follows the customer, so this usually traces to a short or mixed opening message that was genuinely ambiguous.

We can bias it toward a default language for your market if that suits you better than pure detection.

Contact us through bobwork.ai/support with the conversation in question. Sending the actual conversation rather than a description is the single thing that speeds up a fix most.

If it is commercially urgent — a wrong price quoted to a real customer, a channel down — say so, and it gets treated as urgent rather than routine.