
After forty-plus discovery calls with controllers, CFOs and shared-service leads this year, one pattern holds across every geography and ERP. Finance teams ask us for one of two things, and the ask depends less on company size than on how ready their data and processes are.
The first ask sounds like this: "We moved to NetSuite last quarter. Give us something simple. Match the bank to the ledger, tie the GL to the trial balance, stop the spreadsheet from breaking." The team wants reliable, repeatable automation with no judgment inside it.
The second ask sounds different: "Our accrual process runs 600 to 700 lines a month, four people, three days. Email the PO owners, read their replies, propose the number, draft the journal. We will review the exceptions." The team wants an agent to execute a judgment-heavy task end to end, with a reasoning trail they take to the auditor.
Internally we call these foundational automation and autonomous execution. Same platform, same Noa agents, different depth of work. The rest of this post lays out what decides which one fits, and what we learned from the teams who asked for each.
Every call we run ends up scoring the same two things, whether or not anyone says so out loud.
Data readiness is about access and connectedness. Does the team have a live trial balance feed, or a Crystal report someone exports on day two? Are the subledgers reachable (AP, open POs, vendor master, bank feeds, PO system), or does the close tool only ever see GL balances? Is the ERP stable, or mid-migration with a chart of accounts being redesigned? One controller at a fintech lender told us her team could not say whether the data coming out of their expense tool and ERP was clean enough for an agent to rely on. Honest, and common.
Process readiness is about whether the work is defined and repeatable. Is there one accrual method, or ten people with ten templates? Are preparers and approvers assigned, or does the checklist live in a spreadsheet with a tab per person? Does the team know which five tasks eat the most hours? A consulting group with teams across four regions described their close as "10 people, 10 different approaches." A staffing company with 41 entities reconciles intercompany through pivot tables, and nobody owns the template.
| Factor | What we check on the call | Low signal | High signal |
| Data | Trial balance feed | Manual export, refreshed daily at best | API or scheduled SFTP, refreshed multiple times a day during close |
| Data | Subledger reach | GL balances only | Open POs, AP postings, vendor master, bank feeds, PO system all reachable |
| Data | ERP state | Migration in flight, chart of accounts changing | Stable for two or more close cycles |
| Process | Task definition | Excel checklist, tab per person | Named preparers and approvers per task and entity |
| Process | Method consistency | Each entity or person has own template | One method per task, exceptions documented |
| Process | Effort visibility | Everything takes long | Team names its top three to five time sinks |
The two factors move independently. A company with clean, connected data and no process is a different conversation from a company with a disciplined process running on exported PDFs. Before the matrix, here is where each kind of agent sits across the five stages every close runs through.
Put the two factors on axes and four starting points appear. The top-right is where agents execute end to end. The top-left is where most mid-market teams start, and where a surprising number of large enterprises sit too. The bottom-left is a polite no for now. The bottom-right is a sequencing decision, not a product decision. Place yourself below before reading on.
The ask is for reliability, not intelligence. These teams want the agent to do one deterministic thing every month without a person rebuilding the workbook. Three situations recur in our calls.
The recent or in-flight ERP migration. A UK services business moving from Sage 200 to Intacct asked us to come back in January, after two clean closes on the new system. A meat processor consolidating three Infor products into M3 by 2027 wants to start background work two to three months before go-live, not before. A UK subsidiary of a global HVAC group with 135 legal entities and ERPs ranging from SAP to QuickBooks is replacing a 30-year-old consolidation system first, and everything else waits. In each case the right answer was a checklist with owners, a trial balance feed, and nothing with judgment in it.
The disciplined team whose data arrives as exports. A global meal-kit company runs a five-day close across 35 entities on Google Sheets linked to the GL. Their incumbent close tool ran for three years and never connected to the AP subledger, the asset register or open payables, so every reconciliation meant an Excel download and an upload. Their ask was blunt: tools for small local teams must reduce the burden, not add to it. What fits is scheduled SFTP report drops, rule-based matching on those drops, and Excel builders which regenerate their existing schedules from raw data. The process was already sound. The data path was the gap.
The high-volume matching problem. A Polish payments company reconciles 300 bank accounts, with a single bank rec taking up to 20 days and transactions going missing for weeks between their data gateway and the ERP. A nonprofit events business matches point-of-sale, processor and bank deposits across three time zones for 20,000 attendees. A semiconductor company already gets 80 to 85 percent auto-matching out of Oracle and told us, correctly, a bank rec agent would add nothing. Matching is where rules do the work. Rules alone reach 85 to 90 percent on a first run in our deployments, and the AI layer, which runs only on the unmatched residue, pushes it to 97 to 98 percent by month two.
| Foundational ask | What the agent does | What stays with the team | Typical stand-up |
| GL to trial balance tie-out | Pulls TB on schedule, compares to workings, flags differences above threshold | Reviewing flagged differences | Under two weeks |
| Bank, vendor, intercompany matching | Runs configured rules, clusters probable matches for the residue | Confirming or rejecting suggested matches | Two weeks |
| Schedule automation | Regenerates linked Excel workbooks from the data of the month | Reviewing the reconciling items | Set up once at implementation |
| Checklist with trial balance | Assigns preparers and approvers, tracks status by entity | Doing the tasks | Two to three weeks |
None of this is glamorous. All of it is what the teams in the top-left quadrant asked for, and it is what gets them to the top-right.
The ask here is for the agent to own an outcome. Not "flag the variance" but "email the PO owner, read the reply, propose the accrual, draft the journal, build the schedule, and show me the exceptions."
These teams share a profile. The ERP has been stable for a while. Open POs, AP postings, vendor master and unapproved invoices are reachable through an API or a reliable scheduled pull. Someone on the team has already named the process as the problem. The questions they ask on the call are about judgment and defensibility, not about connectivity.
The accrual-heavy enterprise. A semiconductor manufacturer runs 600 to 700 accrual lines a month off Oracle and Concur. Their bank rec was fine. Their flux took two days. Their accrual process was the fit, and the first thing they wanted was a live run on their own invoices to test the goods-versus-service split. A data analytics company on Coupa and NetSuite maintains accruals in one inception-to-date file where lookups break every month. Their question was sharper: where exactly is the AI doing work, versus applying a rule? The honest answer is both. Pro-rata and historical-average estimates are arithmetic. Reading "95 percent of the service is complete" in an email from a PO owner and turning it into a number with a confidence score is inference.
The team with a strong close looking past it. A cybersecurity company closes in five days with no reopens in a year and told us their accrual process has the fewest misses of anywhere their finance lead has worked. They did not need help closing. They asked whether an agent could compute and forecast free cash flow from collections, disbursements and payroll. When the foundation is this solid, the ask moves beyond the close.
The budget-based accrual. A UK lender raises blanket annual POs for 80 to 85 percent of budget and identifies the rest through business partnering. Their Excel workings became the rule set. The agent emails business owners through a link, reads the reply, and stages the debit and credit from the chart of accounts. Their open question was data quality in the narrations of their PO system, which is a fair worry and the reason the agent shows its reasoning on every line.
| Autonomous ask | Inputs the agent needs | What it produces | What the team reviews |
| Accruals | Open POs, AP postings (3 to 8 months), vendor master, unapproved invoices | Two estimates per line (rule-based and inference-based), confidence score, JEs, schedule by vendor and GL | Exceptions and low-confidence lines |
| Prepaids | AP ledger, prepaid GLs, chart of accounts, vendor whitelist | Prepayment master, amortisation schedule, corrective entries | Newly identified prepaids |
| Intercompany | Ledgers of both entities, FX table | Matched population, residue with suggested matches, elimination entries | Suggested matches above tolerance |
| Flux commentary | GL history, sub-GL, budget, business metrics | First-draft commentary per entity, rolled up to group | Commentary before it reaches the CFO |
In our deployments the accrual cycle drops from two to four days with four to ten people to one to two days, roughly 60 to 70 percent less effort. The number matters less than what the team does with the time. In every case so far they moved to reviewing exceptions and the reasoning trail, which is the role Kaushik described in The Defensibility Problem: the agent prepares, the accountant decides, and the chain of thought is the audit trail.
The question from the top-left quadrant: why buy a platform with agents I will not use for a year? The answer is a mapping, not a pitch. Each ask gets a specific set of agents and components, and the components for the foundational ask are the first half of the components for the autonomous ask. Your answers above select your quadrant below.
Three things make the mapping hold.
One data path. The SFTP drop or API a team sets up for foundational matching is the same feed the accruals agent reads later. The chart of accounts ingested for the trial balance tie-out is the same master the prepaids agent maps journals against. Nothing is rebuilt between the foundational step and the autonomous step. A meal-kit company described their path as manual Excel, then automated Excel, then agents. The path works because each step adds the data access the next step needs.
Two agent tiers, one output. Matching agents are deterministic: every match carries a rule number and a unique code, and the AI layer sees only the unmatched residue. Accrual, prepaid and lease agents add inference on top: every inferred number carries its reasoning and a confidence score. Both tiers produce the same two outputs, journal entries and a reconciliation schedule. A team moving from the first tier to the second changes nothing about how it reviews, posts, or hands work to the auditor. Pricing follows the same split, so a team in the top-left quadrant pays for rules, not for inference it cannot use yet.
One set of guardrails. The controls under a two-entity bank rec are the controls under an enterprise accrual run: calculations in the application layer, the model sees structured outputs rather than raw ledgers, every action logged with a timestamp, read-only auditor access. This is what a consulting firm meant when it rejected building its own agents on a frontier model (no owner for maintenance, no consistency across clients, no audit trail), what a staffing company holding an ISO 42001 certification meant by human-in-the-loop as a condition of keeping it, and what Karthik laid out in The Build vs. Buy Question in Finance Close Has a New Answer. WPP runs the flux module across 3,000 subsidiaries with a Big 4 auditor who has cleared the platform two years running.
The same team runs both phases: professional accountants who code. The forward-deployed accountant who scopes the foundational rollout returns a quarter later to switch on the accruals agent. One diagnostic, two phases, no second discovery.
Two practical notes from the calls. First, ask the vendor to run on your own data before you commit. Every serious evaluation we have run started with a prior quarter of POs and AP postings, and the output compared against what the team posted. Second, pick the process the team hates most, not the one the demo looks best on. A controller who cares about prepaids will not be won over by a bank rec.
Some of the best conversations we had this year ended with us saying "not yet." A mining company with ten consultants on Xero called a full platform a sledgehammer to crack a nut, and for their monthly close they were right. Their quarterly consolidation was a different matter. A group CFO replacing a decades-old consolidation engine told us to come back after wave one. We will.
The pattern underneath all of it: the agent is never the constraint. Data access and process discipline are. Teams with both get execution. Teams with one get a foundation and a path. Teams with neither get a conversation about sequencing, and a promise to return.
Related reading
Based on anonymised Consarc discovery and demo calls, April to October 2026. Company descriptions are altered enough to prevent identification. Metrics on match rates, stand-up times and accrual effort are from Consarc deployments as stated on those calls.