Integrate an AI agent where your users already are
I am Atif Iqbal, and I build AI agents into websites, WhatsApp, Slack, Microsoft Teams and Jira. Not a widget pasted on top: an agent that knows who it is talking to, acts only with their permissions, asks a person before anything it cannot undo, and keeps a record of everything it did. Each channel is quoted as a fixed price against a written scope.
Send the six answers the email template asks for and you get the channels, the weeks, the approval gates and a fixed price back, usually within a day.
Five channels, with the weeks and the gates
The engineering is the part you control. The approval each platform requires is the part you do not, so it goes on the timeline from day one instead of surprising you in week four.
| Channel | What you get | Typical build | The gate you do not control |
|---|---|---|---|
| Your website | An agent inside your site or product that answers from your own data and acts for signed-in users, with hard caps for anonymous visitors. | 2 to 4 weeks | None. You ship when it is ready, which is why it usually goes first. |
| Customers message your business number and the agent answers, books and updates, with account linking rather than trusting a phone number. | 2 to 3 weeks | Meta business verification, then review of every template the agent sends first. | |
| Slack | An agent your team mentions in channels and threads, on the narrowest scopes that do the job, replying in the right place. | 1 to 2 weeks | A workspace admin installing it. |
| Microsoft Teams | The same agent in Microsoft 365 organisations, with a permission list the customer's administrator will actually approve. | 2 to 3 weeks | The customer's tenant administrator consenting to the permissions. |
| Jira | An agent that triages, comments and moves work using each user's own Jira permissions, without a vendor holding your credentials. | 1 to 2 weeks | A project admin, plus Atlassian review if it is distributed. |
The weeks assume the agent core exists: the tools, identity mapping, state and audit trail every channel shares. The core is scoped in the first conversation, because its size depends on how many of your systems the agent has to act in. If you are not sure which channel comes first, where your agent should live walks through the decision, and integration by channel compares the engineering for each.
What makes it safe to switch on
A demo agent answers well. A production agent also knows who it is talking to, survives retries, and stops at the edge of what it may do alone. These four are in every quote, not offered as extras.
Identity before any tool runs
A Slack id, a phone number or a cookie is not a user. Each one is mapped to an account in your system, and the agent only sees what that person could.
How it is doneRetries that do not double anything
Every platform redelivers events. Inbound messages and tool calls are idempotent, so a retry never sends a second reply or issues a second refund.
How it is doneA person where the risk is
Bounded loops with step, time and cost limits, and explicit approval for anything that moves money, deletes data or grants access.
How it is doneAn audit trail and evals
What was asked, which tools ran, as whom and what came back, plus a test set of real questions the agent has to keep passing.
How it is doneHow the price is set
One fixed figure for the core and each channel, split into milestones you pay as each one ships, with every platform approval marked on the timeline as the gate it is rather than buried in a contingency. Five things move the number more than anything else:
- 01How many systems the agent has to read from or act in, and how good their APIs are
- 02Whether users are anonymous, signed in, or staff in somebody else's workspace
- 03Which actions it may take on its own and which need a person to approve
- 04How many channels, since the core is built once and each channel is an adapter
- 05Data that must not leave your infrastructure, which decides where the model runs
Sometimes the right answer is not to build. If the agent only answers questions about your public pages, a vendor widget is cheaper and I will say so. Where that line sits is in adding an AI agent to your website, and for support agents in build versus buy.
AI systems already shipped
How the retrieval, agent and model work is built and evaluated is on the AI engineering page.
What teams ask before adding an agent
How do I integrate an AI agent into my website?
If it only answers questions about your public pages, a vendor widget and a script tag is a reasonable answer. Once it needs your customers' own data or has to take actions, build it: a core that holds the tools, identity and audit trail, and a website adapter on top, which takes two to four weeks once the core is scoped.
Which channel should the agent live in first?
Wherever your users already do the work. Customers who message you on WhatsApp will not open a chat widget, and staff who live in Teams will not visit a portal. Build the core once with the first channel, and file the second channel's verification or consent in the same week, because those approvals run on calendars you cannot speed up.
Can one agent run in several channels?
Yes, and that is how I build it. The tools, identity mapping, state and audit trail live in one core, and each channel is a thin adapter. Adding Teams to an agent that already works in Slack is two to three weeks rather than a second project.
Which AI model do you use?
Whichever fits the task, cost and data rules. The core does not depend on one provider, routine work can go to a cheaper or smaller model, and if data cannot leave your network the model runs inside it.
What stops the agent doing something it should not?
It only gets tools for what it is meant to do, it acts as the user and never with more access than they have, every loop has hard limits, and anything irreversible waits for a person to approve. Every step is logged, so you can always answer what it did and for whom.
Who owns the agent and the accounts?
You do. The code, the cloud and model provider accounts, the Meta business account, the Slack and Teams app registrations are all yours, with access granted to me, and ownership is assigned on final payment.
Tell me what the agent has to do and who is allowed to ask it. The channels, the weeks and the price come back with the scope.