Business AI: improve decisions, not just deliverables

Business AI: improve decisions, not just deliverables

The leadership team gets its quarterly review ahead of schedule. The charts look good. So does the commentary. Yet one question remains unanswered: why did it take this meeting to discover that several customer accounts were slipping?

AI improves a decision when it helps people spot a problem, test competing explanations or choose an action using evidence they can check. Producing the deck faster may be valuable. But if the same surprises arrive at the same time, management has not improved yet. Those are two separate benefits to assess.

Consider an illustrative example, not a client case study: a midsize business wants AI to prepare its quarterly business review, or QBR. “Build the slides from our numbers” sounds like one request. It can actually describe several different projects, each needing different data and a different test of success.

Is the document really the problem?

If the team spends its time copying tables into slides, automating that work solves a real problem. Check the numbers, compare total preparation time — including review — and decide whether the result is worth keeping. Not every presentation needs an AI agent programme behind it.

The brief changes when a leader wants to understand falling margins, redirect sales effort or identify a risk before the next meeting. A polished deck is no longer enough. The analysis must connect an observation to a choice: which account needs attention, which explanation needs checking, and who should act?

Before choosing a tool, ask: “If the document were already finished, what would still stop us from making the decision?” That unanswered question sets a useful level of ambition.

From a faster QBR to earlier warning

In our example, the first level is producing the presentation with traceable figures. Success means a reliable review pack that requires less manual preparation.

The next level is preparing the decision. AI brings together revenue changes, renewals and customer feedback, separating facts from possible explanations. A drop might reflect billing timing rather than a departing customer. The sales lead checks that explanation before reassigning the team.

A further step would be detecting warning signs between reviews. Instead of explaining the past quarter, the system flags a change while someone can still intervene. That requires current data, a rule for raising an alert and follow-through on what happens afterwards. A one-off analysis does not demonstrate this capability.

The same gap between a record and its consequences appears in our guide to Twenty CRM and the next sales action: a meeting summary does not fix follow-up unless the next action is assigned and checked.

A more ambitious brief needs a deeper investigation

Asking for “more strategic” analysis does not give a model the evidence it lacks. Reporting commentary already reflects choices about periods, scope and exclusions. To check a variance, the team needs access to the authorised records behind it and shared definitions of the metrics.

For the QBR, that might mean comparing renewal dates, service incidents and sales conversations. It does not mean opening the entire information system. Start with the data needed to test the explanation, within the access rights of the people involved, and make each material claim traceable to its source.

The method matters too. Ask for competing explanations, look for a case that contradicts the conclusion and have the relevant business owner review the interpretation. Finance may explain a billing delay; operations may identify a recurring service issue; sales may know a renewal has already been negotiated. Combining those perspectives helps distinguish a symptom from its cause.

That depth comes at a cost. In Building effective agents, Anthropic recommends starting with the simplest solution and notes that agentic systems often trade latency and cost for better task performance. Added sophistication needs an established purpose, not just an impressive demonstration.

How do you know the decision is getting better?

A model can make a persuasive recommendation and still be wrong. Asking it to approve its own analysis is not independent evidence. Verification must examine the underlying data, known cases and observable consequences of using the system.

Anthropic makes a useful distinction in Demystifying evals for AI agents: what an agent says in a conversation is not necessarily the state it has created in the environment. Saying a flight is booked does not prove that a reservation exists. In our QBR example, a well-written alert likewise does not prove that a genuine risk was detected in time.

A test should therefore examine useful alerts, false alarms and missed problems. It should also account for the checking work imposed on the team. If every alert requires someone to redo the analysis, the benefit may disappear even though the deck takes much less time to produce.

Replaying past periods can reveal errors, provided the system receives only information available at the test date. Otherwise it may “detect” a problem whose eventual outcome it already knows. Follow that initial check with observation alongside the existing process. Neither step alone proves an improvement in revenue or customer retention.

A five-field brief for the next request

Complete these five fields before commissioning the work. They usually reveal whether you are buying a document, decision support or an early-warning system.

• Deliverable: what must be produced? For the QBR, a review pack and a sourced list of variances.
• Expected outcome: which decision should become possible or better informed? For example, selecting accounts for investigation before the review, with a named owner.
• Data: what evidence can verify the variances? Agreed metric definitions, records available at the right date, accessible sources and explicit coverage limits.
• Permissions: what may AI read, propose or change? For the first test: read an authorised dataset and make internal recommendations, with no customer messages or CRM changes without approval.
• Proof: what will justify continuing or stopping? Useful and missed signals, time to identification, review effort and documented decisions, compared with the current method.

The executive sponsor approves the target decision and unacceptable errors: for example, no at-risk account should be dropped from review solely on AI’s recommendation. Business and technical teams define test cases, access and comparison rules together. The leadership team can then judge whether the tool helps, without having to settle every technical detail itself.

Test a bigger ambition before granting more authority

Run the test alongside the usual review preparation. AI produces its analysis without contacting customers or triggering actions. A responsible owner compares its proposals with the team’s work, records disagreements and checks what explains them. Set a review date and continuation conditions before the test begins.

At that review, three choices remain open: keep only the document automation; extend the analysis support if it adds verified evidence; or stop because data quality and checking costs leave no useful gain. Our article on proving value on a small scope before a large AI project explains why a bounded test lets a business make that choice before committing to a wider rollout.

For this kind of project, the Say Digital Framework starts with a specific decision, then moves through a bounded prototype and evidence checks before a deployment decision and ongoing measurement. Verification makes greater ambition possible. It should not be used to justify a tool the business has already decided to buy.

Is your next AI request about reporting, analysis or a leadership review? Bring the question that remains open after the document is finished. Talk to Say Digital about the expected outcome, necessary data and a first test you can verify.

Sources and limitations

Sources checked on 25 October 2026. The references below document design and evaluation principles; they do not establish a commercial gain for the illustrative scenario in this article.

• Anthropic — Building effective agents: simplicity, cost and latency trade-offs, and feedback from the environment.
• Anthropic — Demystifying evals for AI agents: observable outcomes and combining evaluation methods.