How to budget for AI automation in 2027

Field notes. John Arndt, Soxoa. Published November 13, 2026.Written November 2026 as a planning guide for 2027. It will be revised when its facts are re-checked.

Budget AI automation for 2027 one workflow at a time: list the one-time costs to build and test it, the recurring costs to run and review it, and a reserve for running it by hand, then set them against the measured hours the workflow takes today. If the recurring lines would consume most of the capacity the workflow frees, or the work cannot be measured, leave it manual for now.

This guide gives no figure to budget, because no honest figure exists before you have measured your own workflows, and it does not forecast what AI models or tools will cost in 2027. It gives you the cost lines, a way to compare them with the benefit using Soxoa's free planning tools and their stated assumptions, a fictional worked example, and the cases where the right budget line is zero.

Budget by workflow, not by tool

A budget line called AI tools tells you what you spent, not what you got. A line per workflow, with its own costs and its own measured hours, tells you whether each one is worth keeping. Split a tool that serves several workflows by use; the monthly cost guide shows how to work out cost per accepted output.

If you have not chosen the workflows yet, start with the 2027 automation plan. In that plan the first quarter's budget is mostly staff time to measure a baseline.

One-time cost lines

Each line is a cost whether your own team, a contractor or a consultant does the work.

Mapping and baseline

Staff hours to write down how the work runs and to count it over typical weeks. Small, and every other line depends on it.

Build or configuration

Connecting the systems and setting up the automation. The consulting cost guide sets out what moves this scope up or down.

Evaluation

Time to collect approved examples, run them, and have the owner check results against the acceptance bar.

Access, security and privacy review

Approvals for the data the workflow touches and where it is processed. If this review is slow, put it in the plan as a line rather than meet it as a surprise.

Training and handover

Time for the people who will review and own the workflow to learn the new path and the manual fallback.

Recurring cost lines

These continue every month the workflow runs. Budget each one per workflow.

Per-seat subscriptions

Assistant seats or AI add-ons charged per user, due whether or not the seat is used.

Metered usage

Charges per unit of work for model calls and automation runs. They move with volume and with settings, so budget them from a measured run rate, not from a vendor's example.

Riders on tools you already pay for

Price changes on existing plans attributed to AI features. Compare each renewal with last year's invoice.

Review time

The reviewer's minutes on every item that keeps a human check, at their loaded hourly cost. It is easy to leave out because it never arrives as an invoice.

Ownership and upkeep

The owner's time to watch the workflow, handle exceptions, and re-test after a model, vendor, source system or rule change.

Reserve for manual cover

Enough to run the workflow by hand for a period you choose if the automation is paused: the baseline's weekly hours, times loaded cost, times the weeks of cover you want.

The benefit side: capacity, not cash

Soxoa's free ROI calculator estimates the value of repetitive time from three inputs: how many people do the work, hours per person each week, and their fully loaded hourly cost. Its planning view assumes 60% of that time can be covered, across 48 working weeks, with 40% and 80% views beside it to show how much the answer depends on that assumption.

The estimate_automation_roi tool in the Soxoa MCP planning tools does the same arithmetic for an assistant. When you leave out the share and the working weeks it uses 60% and 48, and it calculates payback only when you supply an implementation cost.

Neither is a forecast. The 60% is a planning assumption, not a measured result, and the output is capacity value: hours that become available. Those hours become cash only if they go to other paid work, absorb growth without a hire, or cut overtime or contractor spend. Write down which of those you expect before you count the money.

A worked example: supplier invoices at a small practice

A fictional accounting practice has three staff who each spend 6 hours a week entering supplier invoices, at a loaded cost of $40 an hour. At the calculator's 60% planning assumption, 3 people × 6 hours × 60% is 10.8 hours a week. Over 48 weeks that is about 518 hours a year, worth about $20,736 in capacity.

The practice adds its recurring lines: $250 a month in usage and subscriptions ($3,000 a year), and 2 hours a week of review at $40 across 48 weeks ($3,840 a year). Net capacity value is $20,736 less $6,840, or $13,896 a year.

Suppose the one-time build and evaluation cost $12,000. Against gross capacity value, payback looks like about 7 months. Against the net figure it is about 10.4 months. At the calculator's 40% view, net capacity value falls to $6,984 a year and payback stretches to about 20.6 months.

Same workflow, same costs, and the answer moves from under a year to nearly 21 months on one assumption. So the practice budgets the pilot first, measures the share it actually achieves, and decides the wider rollout from that number. It keeps the 40% case in the budget, not only the planning case.

The practice and figures are invented. They are not an estimate of what any real workflow will save.

What to measure every month in 2027

Measure the same things for every workflow, so the budget can be compared line by line at the end of the year.

Accepted outputs

Finished items that passed review. The unit everything else divides by.

Cost per accepted output

The month's recurring charges, including failed attempts and retries, divided by accepted outputs.

Review minutes per item

If this does not fall as the pilot settles, the saving is smaller than the plan assumed.

Exception rate

Items that left the normal path, as a share of all items. A rising rate is a reason to check whether the inputs changed.

Covered share

The share of the original manual time the workflow now covers, measured. Replace the 60% assumption with it as soon as you have it.

Hours actually redeployed

What the freed time went to. Without this line, capacity value stays a number on a spreadsheet.

When the right budget line is zero

Leave a workflow manual, and spend nothing on automating it this year, in these cases. A leave it manual decision with a date to revisit is a budget decision too: it frees money for the workflow that clears the bar.

It is rare

Low volume means a small saving and too few cases to evaluate.

Review costs as much as doing

If every output needs a full check by the same person, the automation moves the work instead of removing it.

A mistake is irreversible

Money leaves, a filing goes in, a customer is told something wrong. Start somewhere else.

Nobody will own it

Without a named owner, nobody is accountable for the recurring lines.

The recurring lines eat the benefit

If review, usage and upkeep take most of the capacity value at the 40% view, wait or narrow the scope.

Building the 2027 budget: a checklist

1. Measure before you price

   Count runs and minutes for each candidate workflow over typical weeks.

2. Write both cost columns

   One-time and recurring lines for each workflow, including review time and the reserve.

3. Run the numbers at 40% as well as the planning case

   Use the ROI calculator or the MCP tool with your own inputs, and keep the low case in the budget.

4. Say where the hours go

   Name the work the freed time moves to, or say plainly that the value stays as capacity.

5. Fund the pilot before the rollout

   Commit the wider rollout from the pilot's measured share and cost per accepted output.

6. Set the review dates

   A monthly cost check per workflow, and a re-test whenever a model, vendor or rule changes.

Paying for outside help

Soxoa publishes this guide and sells some of the help described here, so weigh this section with that in mind. Self-serve tools may be enough for simple, low-stakes connections (DIY tools comparison), and a hire may be better when the work needs daily judgment (automation vs hiring).

Soxoa's consulting engagements are quoted after a call, the readiness assessment is free, and the agent-ready build for SaaS has a fixed price range published on its own page. What Soxoa costs sets out each one. Whoever you hire, get the scope, the acceptance bar and the price in writing before work starts.

What would change this advice

This guide was written in November 2026 and will be revised when its facts are re-checked. Rebuild the numbers rather than carry them forward if any of these happen.

A vendor changes how it bills

If a tool moves from seats to usage or back, rebuild that workflow's recurring lines from the new terms.

Usage charges move

If model or automation prices change during 2027, recalculate cost per accepted output. This guide assumes nothing about which way they will move.

Volume changes

Growth or a lost customer changes both the cost and the benefit. Measure again.

Freed hours have nowhere to go

If there is no other work for the time, the value is capacity only, and the case has to rest on speed or accuracy instead.

The planning tools change

If Soxoa changes the assumptions in its calculator or MCP tools, this guide will be revised to match. As written they use 60% coverage and 48 working weeks.