The bold claims about AI in finance still outrun the reality in most close cycles. Close has not become autonomous. Controllers are not handing the books to a model and going to the beach. The process is still shaped by deadlines, judgment, policy interpretation, late adjustments, broken data handoffs, and the simple fact that numbers have to stand up under scrutiny.
But that does not mean nothing material has changed.
What has changed is narrower and more commercially useful. AI is taking time out of specific close tasks that used to consume senior attention in the worst possible moments: recon support, workbook inspection, variance explanation drafts, commentary assembly, and evidence gathering across fragmented files.
The right way to read 2026 is not “AI has transformed close.” It is “certain parts of close now move faster, with better first drafts and less manual hunting, when the workflow is tightly controlled.”
Workflow one: reconciliation support is finally practical
The strongest direct example in the scraped database comes from the accounting thread where a user describes Claude in Excel handling bank reconciliations. The workflow is concrete: drop in bank and GL activity, specify the ending bank and GL balance, let the model match transactions, identify differences, and suggest journals that may need to be recorded. The same user notes that after adding one business-specific instruction about batch deposits, the reconciliation improved again.
That is the shape of real AI value in close. Not full automation. Pattern matching plus exception surfacing inside a constrained dataset.
Microsoft has been pushing toward the same operating model in Copilot for Finance and broader finance agents. Its finance materials describe identifying unmatched transactions, surfacing differences, and suggesting next steps inside Excel-connected workflows. That matters because reconciliation is one of the least glamorous but most expensive close bottlenecks. When teams are buried in mismatches, the biggest gain is not fewer humans. It is faster movement to the exceptions that actually require accounting judgment.
Governance matters here. Reconciliation support works only if the model is limited to proposing matches and exceptions, not silently finalising accounting treatment. Review needs to stay with a named owner, and the source data needs to sit in an approved environment. The governance model from Article 1 applies directly: classify the task as medium-risk, qualify the model on historical recs, constrain the environment, and verify all proposed journals before posting.
Workflow two: variance commentary is moving from blank page to supervised draft
The database is stronger on commentary than on almost any other finance use case. AI in Finance’s proof-of-concept piece explicitly uses budget-versus-actual commentary as the starter use case for finance teams because it is bounded, repetitive, and easy to measure. That is the right instinct. Commentary has always been an awkward drain on senior finance time because it mixes pattern recognition, business context, and writing discipline under deadline pressure.
AI is now good enough to take the first pass when the inputs are disciplined. Feed it actuals, budget, prior period, cost-centre logic, and a required output format, and it can draft a usable starting point far faster than a human beginning from zero. AI in Finance’s prompt-engineering guidance makes the rule clear: specificity determines quality. If the model gets the driver tree, the audience, the thresholds, and the source tables, the draft becomes sharper. If it gets vague instructions, the output turns generic fast.
Here is a first-draft commentary prompt that holds to that discipline:
You are drafting month-end variance commentary for the [department] P&L.
Inputs (attached): actuals, budget, and prior period by cost centre.
Materiality: explain any variance over the lower of £50k or 5%.
For each material variance, state the driver, quantify it, and separate
price, volume, and timing effects. Mark anything you cannot evidence from
the numbers as "requires review". Output a table, then a three-bullet
summary for the board pack. Do not infer drivers you cannot support.
The prompt does the mechanical work. The controller still owns the sign-off: every driver is checked against the ledger before it reaches the pack.
This is where ChatGPT, Claude, and Gemini are genuinely useful in close. They do not replace the finance leader’s interpretation. They cut the dead time between “numbers landed” and “narrative exists.” For group reporting teams, that is meaningful.
The governance overlay is non-negotiable. Commentary must remain tied to traceable numbers and defined review thresholds. If a model says payroll drove the variance when the actual driver was timing in revenue recognition, the draft can create more work than it saves. That is why supervised commentary works best where there is already a strong close pack structure.
Workflow three: workbook review and formula inspection are getting faster
One of the more credible operator threads in the database comes from FP&A users describing Claude for Excel as a kind of junior analyst. The tasks mentioned are telling: audit complex workbooks, build formulas, reformat models, inspect existing logic, and help untangle calculation structures. That does not sound dramatic, but in close week it matters a lot.
Spreadsheet risk remains one of the most underpriced problems in finance. A model that can read a workbook, explain what a formula chain is doing, flag suspect logic, and propose cleaner formulas cuts time out of one of the worst close activities: tracing someone else’s model under pressure.
Anthropic’s recent Claude Code and Opus 4.7 material suggests a wider shift toward long-running, tool-using agents that can inspect files, reason across multiple steps, and verify outputs. Even though those releases are positioned around engineering, the pattern is transferable to finance operations. The meaningful capability is not code generation. It is sustained file-level reasoning with iterative checking.
For close teams, that translates into faster workbook diagnostics, more reliable support in preparing recurring reporting packs, and less manual effort in cleaning messy logic before circulation. It also suggests why the best results come from tools that can operate inside the actual file environment, rather than copying snippets into a chat window.
Workflow four: evidence gathering is compressing the ugly middle of close
A large part of close is not accounting judgment. It is evidence assembly. Pulling support from emails, files, worksheets, prior-period notes, policy documents, and side conversations. This is the ugly middle of finance work: necessary, time-consuming, and hard to standardise.
OpenAI’s recent enterprise framing points directly at this shift. It is building toward agents grounded in company context, connected to internal systems, and able to work across files and tools. Google is making a similar play with Gemini across Docs, Sheets, Slides, and Drive, explicitly positioning the system as able to synthesise information across files, email, and the web while keeping data safeguarded.
That matters for close because the delay is often not the arithmetic. It is the hunt. Where is the latest support? Which file carries the final assumption? Who explained the movement last month? Which side note in email explains the operational anomaly?
When AI can retrieve and assemble that context inside governed systems, it shortens one of the least visible parts of close. That gain will not always show up as direct headcount reduction. More often, it shows up as fewer late-night searches, faster issue resolution, and less senior time spent reconstructing context that already exists somewhere in the business.
What AI has not meaningfully touched yet
This is where finance leaders need to stay sober. AI has not meaningfully solved final accounting judgment. It has not solved policy interpretation in ambiguous cases. It has not solved ownership gaps between finance and operations. It has not solved poor chart-of-accounts design, bad ERP hygiene, weak close calendars, or missing reconciliations. It has not made broken data architecture disappear.
The accounting discussion in the database makes that plain. Some users feel real gains. Others still see nothing but hype because their core grind is manual extraction, copying, and system fragmentation. They are not wrong. AI layered on top of weak process design only goes so far.
It also has not removed the need for control. If anything, the faster the first draft becomes, the more important disciplined review gets. That is especially true for accrual reasoning, unusual transactions, and anything likely to be challenged by auditors, investors, or lenders.
The pattern finance leaders should act on
The practical takeaway is simple. Treat close as a portfolio of micro-workflows, not as one monolithic process. Pick the tasks that are repetitive, text-heavy, file-heavy, or pattern-heavy. Keep humans on approval, accounting treatment, and exception handling. Measure time removed from bottlenecks rather than dreaming about lights-out close. The same logic drives the broader finance director's deployment framework: the unit of deployment is the workflow, not the team.
That is where the real gains are appearing in 2026. Not in replacing controllers. In giving them fewer mechanical tasks to fight through during the most time-sensitive week of the month.
That is also where the unfair advantage sits. A finance team that can move from rec noise to true exceptions faster, draft narrative faster, inspect workbooks faster, and gather evidence faster will close with more control and less strain than a team still doing all of that manually.
Close is not becoming magical. It is becoming tighter in specific places. For operators who care about reliability, that is a much more valuable change.
Start with The AI Playbook. Free, 45 minutes, written for senior professionals in finance, property, and hospitality.
