Fortune ran a piece recently on why every company suddenly wants an AI model router. The idea is simple once you strip out the jargon. Instead of sending every task to the same expensive AI model, you route the easy stuff to a cheap, fast model and save the powerful, expensive one for the work that actually needs it. Most of what your firm does every day is the easy stuff.

If you run a firm doing $1M to $25M in revenue, this matters more than it sounds like it should. You’re not choosing between AI and no AI anymore. You’re choosing between paying for a Ferrari to drive to the mailbox every single time, or building a system that knows the difference between a mailbox run and a cross-country trip.

What a model router actually does

A model router sits between the work and the AI models doing the work. It looks at each task and decides which model handles it. Reconciling a bank feed against a ledger? That’s routine pattern matching. A cheap, fast model handles it fine. Interpreting a multi-state pass-through entity’s tax position after a mid-year ownership change? That needs a stronger model, more context, and more careful reasoning.

Firms that treat every task like it needs the expensive model end up with automation that costs too much to scale. Firms that route intelligently see their per-task cost drop 60-80% as volume grows, because the bulk of accounting work is repetitive by nature. Reconciliations, data entry, variance flags, first-pass categorization. None of it needs the heaviest model on the market. It needs consistency and speed.

This is the part most vendors selling “AI for accounting” skip over. They sell you one model, one price, and hope you don’t notice the invoice creeping up as you add clients. A router-based approach flips that. Your cost curve should get better as you scale, not worse.

Where the money actually leaks

Three patterns show up again and again when we look inside firms this size.

Month-end and year-end crunch. Somewhere between 30% and 50% of total staff time gets concentrated into about four weeks a year. Everyone is pulling bank feeds, chasing AP and AR balances, checking payroll, and building close packs at 9pm. It burns people out and it pushes advisory conversations to “next month,” which somehow never arrives.

Client onboarding drag. New clients wait weeks for document collection, chart-of-accounts setup, and historical clean-up to finish. We typically see 20% to 30% of new clients delay a full quarter of billable work simply because onboarding stalls. That’s not a small leak. That’s a quarter of revenue sitting on the table for every fifth client you sign.

Advisory time crowded out. Advisory work bills at 2-3x the rate of compliance work, and it’s the relationship that keeps clients for a decade instead of two years. But compliance is loud and urgent, so it always wins the calendar. The advisory conversation that could have happened in March happens never.

None of these are staffing problems in the traditional sense. They’re routing problems. The wrong kind of effort is going to the wrong kind of task, over and over, at scale.

What this looks like end to end

Take month-end close. Right now it probably looks like this: someone pulls bank statements, someone else exports AP and AR, someone checks payroll totals, and then a senior person spends hours reconciling discrepancies and writing journal entries before a partner reviews the whole thing.

A Month-End Close Agent built the right way pulls bank, AP, AR, and payroll feeds automatically, reconciles them, and flags the variances that actually need a human eye. It drafts the journal entries. It builds a partner-ready close pack before your team has finished their coffee. The routine reconciliation work, the stuff that’s 80% of the volume, runs through a fast, cheap model because it’s pattern matching against known rules. The variance that looks off because of a one-time adjustment gets escalated to a stronger model, or to a person, because that’s where judgment actually matters.

That’s the router logic in practice. Not “AI does everything” and not “AI does nothing.” AI does the repetitive 80%, cheaply, and reserves the expensive reasoning for the 20% that deserves it.

Onboarding works the same way. A Client Onboarding Agent runs a guided workflow that collects documents, sets up the chart of accounts, and produces a clean opening trial balance without a partner touching a single spreadsheet until the review stage. Document collection and chart-of-accounts setup are mechanical tasks. They don’t need your most expensive model or your most expensive staff member. They need a system that doesn’t forget to follow up and doesn’t lose a document in an email thread.

An Advisory Insights Agent works differently again. It reads each client’s monthly numbers, surfaces three things worth discussing, and drafts the partner’s talking points before the meeting. This one leans on stronger reasoning more often, because spotting the right three things in a client’s numbers is closer to judgment than pattern matching. Even here though, the router matters. The agent doesn’t need the top-tier model to pull the numbers together. It needs it for the last mile, where the insight actually gets shaped into something a partner can say out loud in a meeting.

If you want a closer look at how ops-side agents like these get built for a firm your size, /omni/ops walks through the build patterns we use, and /omni/advisory covers how the insight and talking-points layer gets structured so partners aren’t starting from a blank page.

The dollar math for your firm

Let’s put real numbers next to this instead of talking in the abstract.

If your firm does $5M in revenue and 40% of staff hours concentrate into month-end and year-end crunches, you’re probably paying overtime or temp staffing to cover it, on top of the burnout cost that shows up later as turnover. If onboarding delays a quarter of billable work for a quarter of your new clients, and your average new client is worth $15,000 to $40,000 a year, that delay alone is a five or six-figure drag depending on how many clients you bring on.

Add in the advisory work that never happens because compliance always wins, and you’re looking at the $60,000 to $180,000 range we see across firms this size. That’s not a hypothetical. That’s what shows up when we actually walk through a firm’s calendar, headcount, and client mix.

A model router doesn’t fix all of that by itself. What it does is make the automation that fixes it affordable at your scale. Without smart routing, running AI across close, onboarding, and advisory prep gets expensive fast, because every task goes through the priciest model available. With it, the cost curve bends the right direction as you add clients instead of the wrong one.

This is worth reading more on if you’re new to how these systems get scoped for a firm. Our /resources/guides section has a few practical breakdowns of what “AI for accounting” actually means once you get past the sales pitch, and /resources/insights has more on how firms in this revenue band are thinking about the next two years.

Where most firms get stuck

The honest failure mode isn’t “we didn’t try AI.” It’s “we bought a tool that does one thing well and called it a strategy.” A chatbot that answers client emails is nice. It doesn’t touch the month-end crunch. A reconciliation tool that only reconciles doesn’t onboard clients faster or free up a partner’s calendar for advisory work.

The firms getting real value out of this are the ones treating it as a system, not a purchase. They’ve mapped which tasks are routine enough for a cheap model, which need escalation, and where a human still has to sign off no matter what. That mapping is the actual work. The AI part is almost the easy half once you know where to point it.

If you want a starting point for that mapping exercise specifically for month-end, we put together a practical worksheet called the Month-End AI Close Map for Accounting Firms. It walks through which parts of your close are routine enough to route to a cheaper model right away and which parts still need a partner’s eyes. You can grab the direct version here if you’d rather skip straight to the checklist.

What an Omni Audit actually gives you

We don’t sell a generic AI package and hope it fits. The Omni Audit is 60 minutes, three outputs, no deck. We look at your actual close process, your actual onboarding flow, and your actual advisory calendar, and we tell you three things: where the leakage is, roughly what it’s costing you, and which of it a routed AI workflow could realistically handle in the next 90 days.

No slideware. No generic ROI slide pulled from a template. Just your numbers, your workflow, and a clear answer to whether this is worth doing now or worth waiting on.

If you’re weighing this against other priorities, it helps to see what the audit actually covers before you commit an hour to it. See Omni for accounting and bookkeeping for a rundown of what we look at and how we scope the recommendation to firms your size specifically, rather than a generic mid-market playbook that doesn’t account for how compliance-heavy your calendar already is.

The firms that get the most out of this tend to book the call before month-end, not during it. You’ll have more attention to give it, and you’ll see the crunch coming with time to actually change something before it hits.

Book a 60-min Omni Audit and bring your last close calendar with you. We’ll use it as the working example.

The next 90 days

Model routing isn’t a future technology you’ll get to eventually. It’s already how the firms ahead of you are keeping automation costs sane as they add clients. The ones still running every task through a single expensive model are going to hit a ceiling on what they can afford to automate, right around the time their competitors don’t have that ceiling anymore.

Start with the close. It’s the most concentrated pain point, the easiest to measure, and the fastest to show a return. A Month-End Close Agent built with proper routing can be live inside a quarter, not a year. Onboarding and advisory prep follow naturally once the close process proves the model works.

If you want to see what this looks like against your actual numbers rather than a hypothetical, the AI audit for accounting and bookkeeping is the fastest way to find out. And if you’d rather browse examples first, our /resources/blog has more detail on how specific firms have sequenced this kind of rollout without disrupting a busy season already in motion.

Sixty minutes. Three outputs. No deck. Book my Omni Audit and let’s find out what your version of that $60,000 to $180,000 actually looks like.