“We need an AI agent to find new customers.” But what exactly are we asking it to do? And how will we know whether it is doing a good job?

Without clear commercial choices, we risk building an expensive machine for producing activity: more prospects, more emails and more CRM entries. But not necessarily more business.

The opportunity is to connect commercial intelligence, decisions and execution into a system that learns where profitable growth comes from.

Six connected stages, one commercial system

Consider an equipment-monitoring software supplier moving into mid-sized manufacturers in France and Germany. Sales-engineering support is expensive. The challenge is to find customers it can serve profitably, with an offer they will buy and adopt.

  1. Market intelligence. Choose opportunities for demand, buying readiness and cost to serve. The largest factories may already have established systems; smaller manufacturers may have unmet needs but face integration or price barriers.
  2. Customer intelligence. Separate verified facts from commercial hypotheses. A new factory is a fact; an urgent need for maintenance software is an assumption to validate.
  3. Channel intelligence. Decide how customers will buy. Local maintenance partners or customer introductions may earn more trust than direct outreach. Test capability, margin and support requirements.
  4. Content intelligence. Give the buying group relevant evidence. Operations needs a downtime case; finance needs payback assumptions; IT needs integration requirements. If full deployment is difficult to buy, test a bounded paid assessment.
  5. Engagement preparation. Qualify before allocating engineering time: confirm the problem, decision owners, data access and willingness to fund an assessment. People approve consequential communications and commitments.
  6. Execution and learning. Follow deployment, adoption, renewal and expansion. Feed customer outcomes, losses, delivery effort and contribution back into targeting, channels, offers and pricing.

If service costs overwhelm margin, winning more similar customers makes the problem bigger. Learning from delivery belongs inside the commercial system.

Define the commissioning brief

Before building, agree eight things:

  • Business outcome: the measurable commercial improvement sought.
  • Scope: one offer, market, segment and workflow to test first.
  • Information: trustworthy, current and lawful sources.
  • Authority: what the system may research, recommend, update or send.
  • Reliability: how to assess accuracy, relevance, missed opportunities and uncertainty.
  • Economics: the cost of building, integration, operation and supervision.
  • Ownership: accountability for commercial value and operational control.
  • Improvement: evidence for scaling, redesigning or stopping.

Include permissions, activity logs, a stop mechanism and tests with ordinary, ambiguous and misleading inputs. Three hand-picked demonstration accounts do not prove that a system will work across a market.

Make the economics earn the investment

The article’s illustrative 100-account monthly workflow costs €6,300 today. AI-assisted operating costs of €3,600 plus €500 in monthly setup allocation bring the comparison to €4,100: €2,200 in potential monthly economic value.

That represents released capacity, whose value depends on productive reuse. It is not guaranteed cash savings or incremental profit, and assumes no conversion improvement. Doubling review time from 15 to 30 minutes per account reduces that potential value to €700.

Where do agents actually help?

Use ordinary automation where the sequence and rules are clear. Use AI analysis where interpretation helps. An agent becomes relevant when it needs to choose successive actions within agreed limits: for example, finding missing evidence in an account brief before submitting it for human review.

That flexibility must justify its additional cost and supervision. The architecture should follow the business process.

Prepare for the customer’s AI, too

A buyer’s assistant needs clear offers, current proof, explicit constraints and accessible information to assess fit. These help evaluation; they do not guarantee a recommendation. Authorised information exchange needs permissions and integration. Autonomous negotiation for complex industrial purchases remains an emerging scenario, rather than a business-case assumption.

Read the complete argument, illustrations and sources on LinkedIn

Related: Zero-Base AI: What business would you build today?