Trusted Business Numbers in the Age of AI

An illustrative example, not a client case.

A distributor’s sales report says gross margin on its top product line held steady last month. The stock report says the warehouse wrote off more than usual. Finance has a third number. By Wednesday, three people have spent two days reconciling spreadsheets, and the buying decision that depended on the answer has already been made on gut feel.

Now add AI. Someone types “What was our margin on product line X last month?” into a chat tool and gets a confident answer in four seconds. Which of the three numbers did it use? Nobody knows.

That is the real opportunity and the real risk. AI can make business answers far faster to reach. Whether those answers are right still depends on the same foundations it always has.

Most established, growing businesses reach a point where operational complexity outpaces their reporting. Typical signals:

• Inconsistent numbers. Sales, finance and operations bring different figures to the same meeting.
• Slow answers. Simple questions take days because someone has to pull data from several systems and stitch it together.
• Growing complexity. More branches, more products, more systems, and the spreadsheets that used to work no longer keep up.

AI raises the urgency. A tool that answers instantly is valuable when the underlying data is sound, and it is a fast way to spread a wrong number when it is not.

Trusted numbers are not a promise that every system always agrees. They mean:

• Agreed meanings. The business has decided what “margin”, “on-time delivery” or “active customer” means, and everyone uses the same definition.
• Traceable origins. You can see where a number came from.
• Appropriate checks. Key figures are tested against reference answers the business already trusts.
• Explainable differences. When two reports differ, there is a known reason.

The payoff is practical: less time reconciling, earlier visibility of problems, better decisions, and AI answers that can actually be tested.

Conversational AI tools such as Claude Dashboards can now turn a plain-English question into a chart or a table. Used well, that puts answers in front of owners and managers who would otherwise wait for an analyst.

The limits are just as real. An AI tool will happily answer from the wrong table, apply the wrong definition, or fill a gap with a plausible guess. So the safeguards matter:

• Give the AI an agreed set of definitions to work from, not raw tables.
• Control what data it can see, beneath the AI rather than relying on the AI to behave.
• Test its answers against reference questions where the right answer is already known. I ran exactly this exam on Claude Dashboards: 17 questions, answers locked in advance.
• Keep a person accountable for decisions that matter.

Testing provides evidence that the tested questions are answered correctly. It is not a guarantee that every possible question will be. That distinction matters, and it should be stated plainly.

You do not need a full data warehouse before you can use AI for business answers. You need to start with one valuable business question and work back:

1. What does the business need to know?
2. What data, definitions and access does a trustworthy answer require?
3. Which critical gaps must be fixed before anyone relies on the answer for a consequential decision?

Improve only the foundations needed for a trustworthy result. For some businesses that is a small piece of work; for others it reveals deeper issues worth addressing first. Either way, nothing is rebuilt unnecessarily.

Readiness assessment. For businesses that are not sure where they stand. You receive a business-led review of the question you want answered, technical findings on your data and systems, a prioritised list of what needs fixing, and a clear recommendation on whether to proceed to a proof of value.

Proof of value. For businesses with a clear question and accessible data. Together we agree the business question and a set of reference tests, then build a working AI analytics experience against your data. You receive the working experience, the documented test results, the limitations we found, and a go, refine or stop recommendation. When the use case is clear and the data and access are ready, a focused proof of value can take one to two weeks; anything larger is scoped separately.

A failed test is not a failed engagement. It is useful evidence. If the results say stop, or fix something first, that is what the recommendation will say. The aim is an informed decision, not a forced larger project.

What comes next depends on what the findings show:

• A defined project to fix the gaps the assessment or proof of value uncovered.
• Fractional data leadership for ongoing senior ownership of your data foundations, without hiring a full-time Head of Data.
• Trusted Numbers Assurance, an optional service for clients with a working solution who want ongoing confidence. Data, systems and business rules change over time, so the reference tests are re-run regularly and any drift is reported and explained. It is a check, not a guarantee, and it is never a compulsory subscription.

Three models, chosen to fit the work rather than the other way round:

• Fixed-scope projects when the outcome can be clearly defined up front, such as a readiness assessment, a proof of value or a specific remediation.
• Specialist days when you need senior expertise for a focused problem: a data model review, a pipeline fix, cloud cost or query performance.
• Fractional retainers when you need continuing senior ownership at a part-time cadence that can scale up or down.

In each case the work, responsibilities and expectations are agreed before it starts. Pricing depends on scope and capacity and is discussed once the need is clear.

FactSmith works best with South African distributors and wholesalers, multi-site retail and manufacturing businesses, where stock, margin and cross-system data problems are the daily reality.

I am Emilio Martin, the founder. I have spent 25 years building data foundations in banking, FMCG, retail, manufacturing, mining and education. The industries differ; the underlying problem rarely does: getting data from several systems to tell one consistent, trustworthy story. That experience shows up as practical judgement: knowing which technical choices matter for the business, which ones do not, and where shortcuts will cost you later.

You work directly with me, from diagnosis through delivery and testing. No hand-off to a junior team. Technology choices are led by your business need: existing systems are reused where sensible, and AI providers are assessed for suitability against your specific use case rather than assumed to be interchangeable.

Results are shared work. FactSmith takes responsibility for the technical delivery and the testing; your team brings the business context and agrees the definitions; and AI supports decisions rather than replacing the people accountable for them.

The first conversation needs no sensitive data. If we proceed, access is authorised by you and limited to what the use case needs. The terms of any AI provider or system involved are checked for the specific use case, and detailed controls are agreed during scoping rather than assumed.

Sometimes the honest answer is not to proceed yet. The question may need sharpening first, a critical data gap may need fixing before AI adds value, a simpler tool may do the job, or the value may not justify the effort. If that is the case, you will hear it.

A disciplined, transparent process: agreed expectations, testing against reference answers, and honest reporting of what works and what does not. Not perfect AI answers, and not guaranteed commercial results.

Claude Dashboards: a 17-question test · All insights · Where data becomes fact