By Dylan Harrocks , Founder of PMI Stack · Published 2 July 2026 · 9 min read
Every finance tool now has “AI” on the box, and for a company built by acquisition the pitch is seductive: point it at your entities and let it produce the group numbers. The reality is more useful and more careful than the marketing. AI genuinely speeds up a group close, and it genuinely cannot be trusted to invent the figures a board relies on.
This piece is for the CFO or operator of an acquisitive group deciding how much to let AI touch the consolidated numbers. If you run several acquired companies on different systems and the month-end pull is a strain, here is what AI can and cannot be trusted to do, and how to get the speed without putting a hallucinated figure in front of your board.
What is AI consolidated financial reporting? AI consolidated financial reporting is the use of machine learning and large language models to speed up combining several companies into one group financial picture, and to let people ask questions of that picture in plain language. It sits on top of the same consolidation work a group always had to do: mapping each entity's chart of accounts to a common structure, eliminating intercompany activity, translating currencies, and building month-end snapshots.
AI does not remove that work. What it changes is the manual grind around it (matching transactions, coding entries, flagging anomalies, drafting commentary) and the interface, so a group number becomes something you can interrogate by typing a question rather than building a pivot table.
The important distinction, and the one this whole piece turns on, is between AI that helps produce the numbers and AI that lets you ask about numbers produced by trusted logic. The first is an accelerator. The second is only safe when the figures do not come from the model's imagination.
What can AI actually do in a group month-end close? It can take real days out of the close and shift skilled people off mechanical work, which is a large prize for an acquisitive group closing several sets of books at once. The evidence is now concrete rather than promissory.
A 2025 MIT and Stanford study found that accountants using generative AI cut 7.5 days off the time needed to complete a monthly close , while improving the level of detail in their reports by 12% and shifting 8.5% of their time from routine processing to higher-value work. For a group folding three or four ledgers together every month, days off the close is the difference between reporting while the numbers still matter and reporting a fortnight late.
The tasks AI is genuinely good at in the close are the repetitive, rules-based ones: reconciling accounts, coding transactions, spotting anomalies against prior periods, and drafting the first pass of variance commentary. A category of AI-native consolidation tools has grown up around exactly this. HighRadius markets a fleet of record-to-report agents that automate a majority of close tasks; Nominal positions itself as an AI-native consolidation platform for multi-entity finance teams; tools like Docyt push AI at per-entity bookkeeping that rolls up to consolidated reports.
The pattern across all of them is the same: automate the mechanical, keep a human on the judgment. None of them, honestly described, hand the whole group close to a model and walk away. That is not a limitation of any one vendor; it is the nature of the problem.
Why don't finance teams trust AI with the group numbers? Because they have watched it be confidently wrong, and a wrong number in a board pack is a career problem, not a rounding error. The trust gap is not vibes; it is measured, and it is wide.
In a Wakefield Research survey of mid-market CFOs released in January 2026, only 14% said they completely trust AI to deliver accurate accounting data on its own . In the same study, 86% said their finance team had already hit at least one instance of inaccurate or hallucinated data from an AI tool, and 97% said human oversight is critical to accuracy when AI is deployed.
Read those three numbers together and the message is clear. Finance leaders are not refusing AI, they are refusing unsupervised AI on the numbers that carry consequences.
A language model is built to produce a plausible-sounding answer, and a plausible-sounding revenue figure that is off by 4% is worse than useless, because it looks right. When the same figure feeds a covenant certificate or a valuation, “looks right” is the trap.
This is the exact fear a group operator brings to any AI reporting pitch: not “will it be fast”, but “will it make up a number in front of my investors”. A serious answer has to solve for that first, before it talks about speed.
How do you make AI numbers you can put in front of a board? You stop the model from producing the numbers at all. The figures come from trusted, typed queries against your data; the AI's job is to pick the right query, read back what it returns, and never do the arithmetic itself. That architectural choice, not the model's goodwill, is what makes an AI answer safe to repeat to a board.
The principle is worth stating plainly, because it is the whole game and most tools blur it:
A trustworthy AI reporting layer treats the language model as a router and a narrator, never as a calculator. When you ask “how is acquisition three doing against its underwriting”, the model does not compute the answer from memory. It selects a predefined query function, that function runs against your consolidated data and returns a structured, pre-computed result, and the model reads that result back to you word for word. Ratios and variances are calculated by the query, not the model, so there is no arithmetic for it to get wrong. If no query can answer the question, the honest system says “I do not have that”, rather than improvising. The number you repeat to your board is therefore the same number that lives in your warehouse, traceable to source, every time. The tell of a system built this way is that it can show its working. Ask it something outside its data and it declines instead of guessing; ask it a real question and every figure traces back to a specific query result you can open. That is what closes the gap between the 86% who have seen AI hallucinate and a number a CFO is willing to sign.
Where does AI help an acquisitive group specifically? In the two places a group has that a single company does not: many source systems to reconcile, and many acquisitions to judge against the plan they were bought on. Both are where manual consolidation hurts most, and both are where a good AI layer earns its place.
Every acquired company tends to keep its own accounting system and its own chart of accounts, so before any group figure is real those have to be mapped to a common structure. That chart-of-accounts mapping and the intercompany and currency work underneath it is exactly the rules-based grind AI accelerates. It is also why spreadsheet consolidation breaks at scale : the manual bridge that held for three entities quietly fails when the fourth and fifth arrive, usually the week the pack is due.
AI does not change what this work is, only how fast it runs. The mechanics of post-acquisition reporting consolidation still have to be right first; AI sits on top of a sound consolidation, it does not substitute for one. Point a model at three sets of books that have not been reconciled and you get a fast, confident, wrong answer.
The second place is the one no generic close tool touches. An acquisitive group does not only need “what are the group numbers”, it needs “how is each deal performing against what we underwrote”.
When the reporting layer holds both the actuals and the original thesis, a plain-language question like “which acquisition is behind plan on EBITDA” becomes a ten-second answer instead of an afternoon in a data room. Operators building AI-native groups, like Vadim Rogovskiy on the RolyPoly podcast , describe this shift from reporting-as-assembly to reporting-as-conversation as the real payoff of AI.
None of this replaces the finance judgment. It means the group view and each deal's performance both come off one trusted source, and the questions get answered in seconds rather than assembled by hand.
What does the PMI Stack dashboard's AI layer actually do? It puts an assistant on top of your consolidated data that answers questions in plain language, with every figure traced to your real numbers, and it lets your team query the same data from the AI tools they already use. We built this into the consolidated reporting dashboard specifically so the AI is safe to point at a board, not just fast.
Ask the group in plain English. An embedded assistant sits alongside the dashboard, so you can ask something like “what is our covenant headroom” or “how is the Madrid acquisition tracking against underwriting” and get an answer drawn from your live consolidated figures. It shows its tool-call state as it works, renders tables and lists in the reply, and streams the answer back. The suggested questions are configured per group, so the assistant opens with the things your board actually asks.
The trust model is the part that matters. Numbers never come from the model's head: every figure is returned by a typed query against your data, ratios and gaps are pre-computed and read back verbatim, and if no query can answer, the assistant says so rather than guessing.
There is a provenance check that flags any figure that cannot be traced. This is the direct answer to “how can I trust AI numbers in front of my investors”: the architecture is the guardrail, not the model's goodwill.
It answers the same across your own tools. Beyond the dashboard, the same data is exposed through a proper connection (an MCP server) so power users can point Claude Desktop or their own tools at the group and ask their own questions, against the same definitions, with the same figures. Every query is logged. You are not locked into one chat box; the group's numbers are queryable wherever your team works.
Two honesty notes, because they are the reason to trust the rest. First, the financial core is the reliable, repeatable part we lead with; CRM and operational reporting are scoped per group where the data supports them, never promised as a universal switch.
Second, you own the warehouse and the model's API key: your data and your auditors, not ours. The dashboard is forked and branded to your group and built on your real data, delivered forward-deployed over weeks.
What can and can't AI be trusted to do in group reporting? Trust it with the mechanical and the conversational; keep a human and a trusted query engine on anything that becomes a signed number. The line is not about how clever the model is, it is about consequence: the further a task sits from a figure someone acts on, the safer AI is to run it unsupervised.
Task Trust AI to… Keep human / trusted logic on… Reconciliation and coding Match and code the bulk automatically, flag exceptions Signing off the exceptions and the policy Anomaly and variance spotting Surface what moved and draft the first commentary Deciding what it means and what to say to the board Answering “what is the number” Route the question and read back a traced figure Producing the figure (a typed query, not the model) The consolidation logic Nothing here COA mapping, intercompany, currency, snapshots The final board figure Present and explain it Ownership, sign-off, and the provenance trail
The through-line is that AI is a fast, tireless layer around a trustworthy set of numbers, not a replacement for the logic that produces them. Get that order right and you keep the 7.5 days AI can save without ever meeting the hallucinated figure that 86% of finance teams have already seen.
If you are running a group on several systems and weighing how much to let AI touch the consolidated numbers, start from the numbers you can trust and let AI accelerate the work around them. The consolidated financial reporting core is the foundation; the choice between buying consolidation software or building a bespoke dashboard is where you decide how that AI consolidated reporting layer actually gets delivered. Either way, the rule holds: let AI narrate and route, never calculate the number your board relies on.
