Which AI agent should your business start with?

AI projects have a funny way of beginning with the tool. Someone sees a demo, everybody becomes excited, and only later does the awkward question arrive: what work was this supposed to improve?

I would reverse that order. You do not need a grand AI strategy before making one part of the business work better. Start with repeated work that gets delayed, missed or corrected—and choose an agent only after the problem is clear.

Begin with the bottleneck

It is normal not to know where to begin. An OECD review of AI adoption by small and medium-sized businesses found that lack of skills and knowledge remains a common barrier. Natural-language tools may be easier to approach, but a useful business workflow still needs clear data, roles and judgement. There is no prompt clever enough to repair a process nobody understands.

Do not ask, “Where can we put AI?” Ask, “What repeatedly makes a customer wait, makes an employee copy information, or makes a manager doubt the answer?”

CHOOSE THE BOTTLENECK, NOT THE BUZZWORD

What repeatedly gets delayed, missed or corrected?

01Customer replies wait too longThe answer exists, but staff keep looking it upCustomer email agent ↘02Meetings create fuzzy follow-upOwners, dates or claims are disputed afterwardsMeeting intelligence agent ↘03Chat orders are copied by handDetails are missed between WhatsApp and the sales systemWhatsApp ordering agent ↘04Updates take hours to assembleThe same reports are opened every morning or weekQuery and briefing agent ↘

↓Start with one focused pilot

Customer email agent

Choose this first when employees repeatedly search for the same order, delivery, account or policy facts before replying to customers.

A good first pilot: one approved inbox, one common question, read-only access to the required records and drafts reviewed by one team. Measure preparation time and how often facts need correction.

Wait if: the approved answer does not exist, the customer record is unreliable, or nobody owns the final reply.

Meeting intelligence agent

Choose this first when meetings end with disputed decisions, unclear owners or reports that repeat numbers nobody checked.

A good first pilot: one recurring internal meeting, visible transcription, a fixed report format and one or two trusted records for factual checks. Measure report-checking time and unresolved discrepancies.

Wait if: participants cannot be properly informed, transcript access and retention are unclear, or nobody who attended will approve the result.

WhatsApp ordering agent

Choose this first when customers already order through chat and employees repeatedly copy product, quantity and delivery information into another system.

A good first pilot: a small product range with stable SKUs, current stock and prices, explicit customer confirmation and human handover for every exception. Measure abandoned enquiries and order corrections—not just message volume.

Wait if: prices change outside the sales system, product names are ambiguous, stock is unreliable or an order cannot be reversed safely.

Business query and briefing agent

Choose this first when a manager or team repeatedly opens the same systems to assemble an update or answer the same operational question.

A good first pilot: one read-only question such as “Which orders due this week are blocked?” or one short scheduled brief. Require named sources, timestamps and visible gaps. Measure preparation time and whether the answer leads to a decision.

Wait if: the question changes every day, teams disagree about the source of truth, or the brief has no clear reader or decision attached to it.

Compare the work before the software

An AI model is only one part of the solution. Compare what starts the job, what information it may use, what output it may create, who remains responsible and how the business will know whether the pilot helped.

COMPARE THE WORK, NOT THE MODEL

Four agents, five practical questions.

Agent	Starts when	Trusted data	Allowed output	Control	First measure
Email	New customer email	Customer, order and policy records	Source-labelled draft	Reviewer sends	Preparation time and correction rate
Meeting	Approved live transcript	Transcript and relevant records	Actions and checked report	Attendee approves	Checking time and unresolved discrepancies
WhatsApp	Customer message or catalogue order	Product, stock, price and delivery data	Recorded order or handover	Rules and customer confirmation	Order exceptions and abandoned enquiries
Briefing	Question or schedule	Approved read-only systems	Sourced answer or brief	Owner makes decisions	Preparation time and briefs used

Use one focused pilot canvas

A narrow pilot is not a small ambition. It is a way to learn without handing a new system broad permissions or asking employees and customers to trust something that has not been tested.

THE FOCUSED PILOT CANVAS

Make the first test small enough to learn from.

01One workflowA repeated job with a clear beginning and end

02One sourceThe record whose facts must win

03One actionExactly what the agent may prepare or complete

04One reviewerThe person responsible for exceptions and quality

05One measureA number or observation compared before and after

AFTER THE PILOTKeep it, change it or stop—with evidence.

Write one sentence for each box before discussing vendors or models. For example:

When an existing customer asks where an order is, the agent may read the order and delivery records and prepare a source-labelled email. A customer-service supervisor reviews the draft. We will compare preparation time and correction rate across twenty suitable enquiries.

That is specific enough to build, observe and improve.

Know when the business is not ready yet

Pause the pilot when any of these are true:

• There is no reliable record whose facts should win.
• The workflow has no clear owner.
• The proposed integration asks for more access than the job needs.
• Nobody has defined when the agent must stop or hand over.
• There is no safe test data or way to reverse an action.
• Success is described only as “using AI” rather than an observable business result.

These are not reasons to abandon the idea. They show what should be clarified first.

Keep the first measures ordinary

Avoid promising a percentage saving before a real pilot. Record a simple baseline, run the workflow on an agreed set of cases and compare:

• minutes spent preparing and checking the work;
• corrections, missing facts and escalations;
• customer enquiries or orders that were abandoned;
• order exceptions and failed write-backs;
• time spent checking meeting reports; and
• whether a scheduled brief was read and used for a decision.

A slower result with fewer dangerous mistakes may be the better result. A faster result that creates extra checking work may not be worth keeping.

Return to the four examples

The first agent should earn the right to become a larger system. If one narrow pilot cannot produce a result people trust, adding more channels, data and automation will not make the underlying problem disappear.

If you want to see each workflow from beginning to end, read part one: beyond polished replies.

Research and helpful links

• Read the OECD review of AI adoption by small and medium-sized businesses
• Use NIST’s Generative AI Profile to understand common risks
• See the NIST AI Risk Management Framework’s human-oversight outcomes
• Review the European Commission’s AI transparency guidance

Research checked on 24 August 2026. Privacy, recording, consumer, employment and AI rules differ by jurisdiction; obtain appropriate local advice before deployment.