If the AI can’t trace its answer to the ledger, it’s a demo

Most of the AI proposals landing on finance directors’ desks this year promise answers. Few said where the answers come from. Hemant Kapadia, chief financial officer of Anaplan, the planning software company, drew the line in an interview with Raconteur published on 22 September: enterprises need answers grounded in their own data that they can act on with confidence, and that is a different thing from the probabilistic response a large language model gives back.

Kapadia told Raconteur that the biggest challenge facing his sector is a belief that AI will one day do every job for everybody. Companies should use it, he said, and many of Anaplan’s product investments are going into it. But companies, he said, still need their data joined up in systems that can carry real operating decisions. His own example was his dashboard: with pipeline, forecasting, workforce planning and expenses in one connected system, he can see what a closed deal or a changed forecast does to revenue, profit and cash almost in real time.

That is the test to put in every AI proposal a UK partner writes. If an answer cannot be traced back to the customer’s system of record, what is being sold is a demo, not a deployment.

Take the view on its merits

Kapadia has a stake in the argument and it is worth saying so. Anaplan’s 30 June announcement of what it calls the Agentic Enterprise promises a “single, auditable source of enterprise truth” and a set of agents for the office of the CFO by October. On 23 September it hired Jared Coyle from SAP as executive vice-president for AI. A CFO praising grounded, connected data is also describing his employer’s product.

His career gives the view weight beyond the pitch, though. Anaplan’s announcement of his appointment, dated 1 December 2022, says he joined from Sirius Computer Solutions, where as chief financial officer he helped more than triple its equity value in about two years, after CFO roles at Mark Andy and SunEdison Semiconductor and senior finance jobs at General Electric and United Technologies. He told Raconteur that GE’s low-margin manufacturing businesses taught him to read the link between operating metrics and the P&L. He has run finance for a large solution provider and for factories where the margin is thin. He knows what an untraceable number costs.

He also told Raconteur that Anaplan had about $600m (£453m) in annual recurring revenue and was unprofitable when it was taken private, and that revenue has since nearly doubled while profitability has improved. Those are the company’s own figures, given in an interview rather than in accounts, so treat them as a claim.

Price in the cancellations

The reason the test matters is the base rate. Research from Gartner, published in June 2025, predicts that more than 40% of agentic AI projects will be canceled by the end of 2027 because costs climb, the business value stays unclear or the risk controls are not there. The same release cites a Gartner poll of 3,412 webinar attendees in January 2025, in which 19% said their organization had made significant investments in agentic AI and 42% had made conservative ones. Gartner also estimates that only about 130 of the thousands of vendors selling agentic AI are real, and it has a name for the rebranding of assistants and chatbots as agents: ‘agent washing’.

For a partner, that is the exposure. The reseller signs the statement of work and the vendor writes the roadmap. When a project joins the 40%, the services margin and the renewal go with it, and the vendor moves on to the next release.

Gartner’s finance practice added a second figure on 24 September. Its survey of 160 senior finance leaders, run from January to April 2026, found that data extraction, accounts payable and receivable automation and report creation generally deliver returns within nine to 10 months, while data management, insight generation and forecasting take longer before they show value. That is worth putting in front of a CFO, because it separates the promises that are cheap to keep from the ones that are routinely oversold.

Put the test in writing

Here is the practical version, in a form that fits on one page of a proposal.

• Trace. For every answer the system gives, name the system of record it came from – the ERP, the CRM, the payroll file – and show the path. If the honest answer is “the model”, say so and price it as an assistant rather than a decision system.
• Recompute. Can a finance analyst reproduce the number by hand from the same source? If the calculation happens inside a language model rather than a deterministic engine, nobody can audit it, and the customer’s auditors will eventually ask.
• Time to value. Set Gartner’s nine to 10 months against each use case you are proposing and say plainly which ones sit in the slower category.
• Exit. State what the customer keeps if the project is one of the 40%: the connected data, the integrations, the cleaned records. If the answer is nothing, the proposal is for a pilot and should be priced as one.

Sell to the sounding board

One more line from the interview. Kapadia told Raconteur that partnership with the chief executive is critical: “A CFO should be in lockstep with the CEO, acting as a sounding board”. Partners tend to sell AI to the CIO or a line-of-business head and treat finance as the final signature. A CFO who sits that close to the CEO is the person who will ask where the number came from, and the answer decides the deal.

Write the proposal for that reader. A deployment leaves them with assets. A demo leaves them with a slide deck.