AI strategy beyond the hype

An AI strategy is a set of decisions

A useful strategy identifies where AI could improve a business task, what must be true for that improvement to occur and who is responsible for the result. A list of tools or demonstrations does not answer those questions. Start with the work and its constraints.

The five stages below are a Bridges planning framework. They connect business diagnosis with data, governance, delivery and evidence. They are intended to structure a discussion with leadership and the people who will operate the system.

1. Define the business question

Describe a task in terms that a process owner recognizes. Which decision is slow, which information is difficult to assemble, and which exceptions require judgment? Name the person accountable for the outcome and the people affected by a change.

• Define the current task, frequency and sources of delay.
• Agree the baseline: effort, quality, service time or another relevant measure.
• Compare AI with process simplification, better search and conventional automation.

For example, a proposal to “introduce AI into procurement” is too broad. A proposal to help a buyer compare submitted supplier documents against an approved checklist gives the team a task, source material and a review responsibility to examine.

2. Examine the foundations

Review the records the system would use, the rights to access them and the interfaces needed to deliver a result. Samples should include missing information, older records and exceptions, not just well-formed examples chosen for a demonstration.

3. Design accountability before autonomy

Define what the system may read, recommend and change. A tool that drafts a comparison is different from one that approves a supplier or issues an order. Limit permissions to the task and make consequential actions reviewable.

NIST’s AI Risk Management Framework provides a voluntary reference for managing risk across the use and evaluation of AI. Apply it in the context of the actual system; citing a framework is not evidence that risks have been addressed.

• Reference: NIST AI Risk Management Framework

• Name the owner who can pause or withdraw the application.
• Record the approval path for actions with material consequences.
• Make uncertainty, evidence and overrides visible to reviewers.

4. Deliver a bounded application

Choose a first release that can be evaluated in the operating environment. Begin with controlled access and representative users. Test incomplete records, conflicting instructions, integration failures and cases where the system should decline to act.

Separate a successful demonstration from operational readiness. The latter also needs monitoring, support, version control and a workable alternative when the system is unavailable. Include the time reviewers spend checking and correcting outputs when measuring performance.

What a field study can tell your investment team

In the November 2023 revision of Generative AI at Work, Brynjolfsson, Li and Raymond studied the introduction of an AI assistant to customer-support staff. They found that productivity effects differed substantially by worker experience, with stronger improvements among less experienced staff and little effect among the most experienced.

• Brynjolfsson, Li and Raymond: Generative AI at Work, NBER working paper

This is a study of a particular support setting, not a forecast for your business. Our implication is to evaluate the people and tasks separately. In the procurement example, experienced buyers may spend time correcting a draft that helps a newer colleague. Compare completed-task quality and total review effort across both groups before deciding whether to expand the application.

5. Scale, revise or stop

Agree the decision criteria before the pilot. Compare observed results with the baseline and review differences across task types. An improvement in straightforward cases can hide additional work or risk in exceptions.

• Assess output quality and the cost of corrections.
• Measure whether staff use the system and where they bypass it.
• Include integration, model usage, evaluation and support in total cost.
• Record unresolved risks and the owner of each action.
• Choose expansion, revision or withdrawal based on the evidence.

An AI roadmap should include foundation work and decisions to defer. That makes it more useful as an investment plan: leadership can see both the opportunities and the conditions needed to pursue them.