Custom AI Agent for Business: The SMB Buyer's Guide

A practical guide to buying a custom AI agent for business: what it does, timeline, costs and compliance notes for SMBs ready to go live.

It's 6:40pm on a Thursday and the inbox has 34 unread messages: two supplier quotes waiting on answers, a client asking "any update?" for the third time this week, and a lead from the contact form that arrived nine hours ago and hasn't been touched. Nobody is being lazy. There just isn't a person whose job it is to sit on top of every channel all day.

That gap has a price. A quote that goes out 48 hours late loses to a competitor who replied in two hours. A support question that sits unanswered turns into a cancelled renewal. A lead that cools before anyone calls it back never becomes revenue at all. None of this shows up as a single dramatic failure — it shows up as a slow leak in the pipeline that's hard to see until someone adds it up at quarter-end.

This guide walks through what a custom AI agent for business actually looks like in production — not a chatbot demo, but a system with tool access, approval steps and an audit trail — using Bharat AI Saathi's SMB AI Ops Agent as the working example, including the offer page with current scope and pricing.

What SMB AI Ops Agent actually does

SMB AI Ops Agent is not a general-purpose assistant. It's built to run one named workflow well, chosen from five fixed playbooks: inbox triage, quote drafting, support deflection, CRM updates, or weekly reporting. You pick the one workflow that's actually bleeding time or money, and the agent is scoped, built and tuned around it.

Underneath, it's a Claude-based agent wired into the tools your team already uses — email, CRM, spreadsheets, calendar — with defined guardrails on what it can do unattended versus what needs a human to approve first. Every action it takes is logged, so if a quote goes out with the wrong number or a CRM field gets updated incorrectly, you can see exactly what happened and why, not guess at a black box.

It's deliberately narrow. A firm running inbox triage doesn't also get quote drafting bundled in — that's a separate scope and a separate build. The narrowness is what makes a 10-day delivery and a fixed price realistic instead of an open-ended consulting engagement.

How it works, step by step

Kickoff call. You describe the workflow, the tools involved, and who currently owns the manual version of the task.
Workflow spec. We document the trigger, the steps, the decision points, and what "success" means in numbers you already track (response time, quotes sent per week, tickets deflected).
Integration build. The agent is connected to the relevant tools — inbox, CRM, sheet, calendar — using existing accounts and API access, not new platforms you have to adopt.
Guardrails and approval steps. We define what the agent can send or update on its own, and what it must queue for a human to approve first.
Eval set and testing. We run the agent against a set of real historical examples from your business before it ever touches a live customer.
Go-live with two-week tuning. The agent runs live, and we adjust prompts, tool logic and approval thresholds based on what actually happens in the first fortnight.
Daily operation under the monthly care plan. The agent keeps running, we monitor it, and one improvement slot per month keeps it aligned as your workflow changes.

Here's what a quote-drafting run typically looks like once it's live:

Trigger: New enquiry email arrives — "Hi, can you quote for 200 branded tote bags, delivery to Manchester by the 30th?" Agent: Extracts item, quantity, spec and deadline. Matches against current pricing sheet. Drafts a quote at £4.20/unit plus delivery, flags that the 30th is a tight turnaround given current production lead time. Agent: Sends draft quote to the sales rep for approval — does not send to the customer directly. Customer: (after rep approves and sends) "That works, can you confirm delivery date in writing?" Agent: Logs the accepted quote to the CRM, drafts a confirmation reply for rep approval, updates the pipeline stage.

Before and after

Situation	Before	After
New enquiry lands after hours	Sits untouched until someone checks email next morning	Drafted response or quote is ready for approval within minutes
Quote turnaround	Rep manually checks pricing sheet, writes email, often next-day	Draft quote generated instantly, rep reviews and sends
Repetitive support questions	Same answers typed out individually each time	Common questions answered directly, edge cases routed to a human
CRM hygiene	Updated in batches, often days late or skipped	Updated as each interaction happens
Weekly reporting	Someone spends an hour pulling numbers into a doc	Report drafted automatically, reviewed and sent
Visibility into what an AI tool did	None — actions aren't logged anywhere	Full audit log of every action taken and every approval given

What it costs and what is included

Item	Price	Notes
Setup (one workflow, one playbook)	$2,500 one-time	Covers spec, build, integrations, guardrails, eval set, two-week tuning
Monthly care plan	$299/month	Monitoring, prompt and tool updates, one improvement slot per month
Third-party platform, telephony, model usage	Billed to you directly	No markup passed through Bharat AI Saathi; you keep your own accounts
Contract	No lock-in	Cancel the care plan at any time; you retain what was built

The payback logic is simple: if the workflow you pick currently costs a few hours a week of staff time or a handful of missed/delayed responses a month, the setup fee is typically recovered within the first one to two months, and the monthly fee is smaller than the labour it replaces or the labour you'd otherwise have to add. We don't publish an ROI multiple because it depends entirely on your current volume and response times.

Delivery timeline

Day 0: Kickoff

Workflow selection, access to relevant tools, and success metrics agreed on the kickoff call.

Days 1–7: Build

Integrations connected, guardrails and approval steps defined, agent built and tested against your eval set of real historical examples.

Days 8–10: Go-live

Agent goes live on the chosen workflow, with a two-week tuning window beginning immediately to adjust behavior based on real traffic.

What you need to have ready

Admin or API access to the tool involved (inbox, CRM, spreadsheet, or calendar) for the chosen playbook
A named person on your side who can approve actions during the tuning period
15–20 recent real examples of the workflow (emails, quotes, tickets) to build the eval set
Sign-off on what the agent is and isn't allowed to send or change without approval
Any compliance or data-handling requirements specific to your industry or region
Confirmation of which third-party platforms you're already paying for (email provider, CRM, telephony) so usage billing routes correctly

Doing it yourself vs done-for-you

	DIY with no-code tools	SMB AI Ops Agent
Time to first working version	Days to weeks of trial and error, often longer	10 days, fixed
Who builds the guardrails and approval logic	You, usually after something goes wrong once	Built in from day one
Audit trail	Rarely set up unless you build it separately	Included by default
Ongoing tuning	Ad hoc, whenever someone notices a problem	One improvement slot monthly, plus monitoring
Cost	Often cheaper up front, but your time isn't free	Fixed $2,500 + $299/month, no lock-in
Best fit	Teams with in-house technical capacity and time to iterate	Teams who want it live and monitored without building it themselves

Both paths are legitimate — if you already have someone technical with time to spare, DIY tools can get you most of the way there. The trade-off is usually time and the guardrail/audit work that's easy to skip until it causes a problem.

Compliance and risk notes

If the workflow touches customer communications, check your jurisdiction's rules on automated messaging and disclosure — some regions require you to disclose when a reply is AI-assisted.
If call recording or transcription is involved in a support or quoting workflow, confirm consent requirements for your state, province or country before enabling it.
Data residency matters if you operate in the EU, UK or similar regimes — confirm where your CRM, email and model provider process and store data.
Review the terms of service for each connected platform (email provider, CRM) to confirm automated access via API is permitted under your current plan.
Keep a human in the loop for anything with legal or financial weight — quotes, refunds, contractual commitments — via the approval step, not full autonomy.
This guide is not legal advice; confirm requirements specific to your industry and location with your own counsel or compliance team.

Mistakes buyers make

Picking a workflow that's too broad. "Automate my inbox" isn't a scope; "triage and route new enquiries" is. Fix: name the single workflow, not the whole department.
Skipping the eval set. Testing on live customers instead of historical examples means the first mistakes happen in public. Fix: insist on testing against real past examples before go-live.
No approval step on anything customer-facing at launch. Full autonomy on day one is how a wrong quote goes out. Fix: start with approval required, loosen it once trust is established.
Treating the care plan as optional. Workflows drift as your business changes; an agent tuned once in month one falls out of step by month four. Fix: use the monthly improvement slot deliberately.
Not having a named approver ready. If nobody's available to approve drafts during the tuning window, the agent's output just piles up unreviewed. Fix: assign this before go-live, not during it.

FAQ

Is this the same as a chatbot on my website?

No. A chatbot typically answers questions in a widget. This agent works inside your operational tools — email, CRM, spreadsheets — to actually draft, route or update things, with approval steps and logging built around it.

Which workflow should I start with?

Pick whichever one currently costs the most staff time or causes the most visible delay — usually inbox triage or quote drafting for sales-driven businesses, support deflection for service-heavy ones.

What happens if the agent gets something wrong?

The audit log shows exactly what it did and why, and approval steps mean nothing customer-facing goes out unreviewed unless you've deliberately loosened that setting after the tuning period.

Can I switch playbooks later?

Yes, though that's typically a new scoped build rather than a free swap, since each playbook has its own integrations and guardrails.

Do I need to buy new software?

No — the agent connects to tools you already use. You keep paying for those platforms directly; there's no new platform lock-in from us.

What if I want to cancel?

The care plan has no lock-in and can be cancelled at any time. You keep what was built during the setup phase.

Next step

Email support@bharataisaathi.com or message us on WhatsApp with the workflow you want to fix and the tools it touches.

Review current scope and pricing on the offer page, and see how engagements run on How We Work.