Build · Essay

Cost per accepted job: where your AI savings went

Greg Moran · November 2026 · 8 min read

Your board wants AI savings. Your P&L doesn't show them. You measure AI by token price and by how many heads it cut. Neither tells you what it costs to get work your customer actually accepts. That number is cost per accepted job. And the biggest drop in it comes when you pair AI with a skilled, AI-enabled global team. The worksheet and 30-day test below let you run it on one workflow.

The AI scoreboard counts the wrong things

Too many AI scoreboards count speed and headcount cut. Too few ask whether the customer's outcome got better. I posted that on X recently, and the data keeps backing it up.

Gartner surveyed 275 service and support leaders this summer. Only 21% said they had reduced human agent headcount because of AI. 7% have already rehired after earlier cuts, and another 11% plan to rehire within 18 months. Gartner has a name for it: the "cut fast, rehire later" trap.

Miro surveyed more than 9,500 leaders and workers (vendor data, and these are self-reported perceptions). 83% said AI improved their personal productivity. Only 56% said the quality of output improved. 59% reported more AI-generated material to review. 55% reported more low-quality AI output that needed human review and rework.

Then there's Jason Lemkin's weekly agent notes. Lemkin founded SaaStr, the large SaaS founder community, and co-founded EchoSign, which Adobe acquired. He moved to cheaper models, and his daily AI spend still went from about $35 to $84. Call volume went up faster than the price per call came down. His AI fact-checker also flagged 18 of 30 accurate notes, and every flag triggered a rewrite. His words: "we paid twice for worse copy." That's one operator's self-report. Still, every founder running agents will recognize it.

Put those together. The first draft got cheaper. The finished work didn't, at least not automatically. The cost moved into review and rework. If your savings math leaves those out, you're reporting a number your P&L will never show.

What cost per accepted job measures

Here's the formula:

Cost per accepted job = (AI and agent spend + team cost to do the work + review and rework time + owner escalation time) ÷ jobs the customer accepted in the same period

A "job" is the unit of work your customer pays for. A resolved support ticket. A release that passed QA. A month-end close that reconciled. A qualified meeting booked.

"Accepted" means the customer, or the owner who answers to the customer, took it without sending it back. Two rules keep the number honest. Failed attempts stay in the cost. Rejected work never enters the count.

AI alone leaves the human layer untouched

AI pushes down the cost of the first pass. Somebody still has to check it and fix what's wrong. Exceptions still need a person.

The biggest companies aren't skipping that step. In a TCS survey of 1,333 leaders at companies with $500M to $20B+ in revenue (vendor data, enterprise sample), only 10% let AI agents execute decisions without human approval across multiple functions. Coinbase CEO Brian Armstrong described his own agent setup this month: lead agents, sub-agents, and one prioritized queue of everything he has to review. "Every change still goes through the same human review and security checks." His stated goal is "fewer corrections per change." More than 14,000 people bookmarked that post.

So the human layer stays. In a lot of companies it grows. And it's often your most expensive people doing it. Senior engineers reviewing agent pull requests. Your head of support reading AI replies at 10pm. That's the line in the worksheet most founders never fill in.

The big win is talent arbitrage plus AI automation

AI lowers the cost of producing the work. Global talent lowers the cost of the people who check and fix the work. Do one and you shrink one part of the formula. Do both and you shrink almost all of it.

Talent arbitrage isn't new. Outsourcers built an industry on it. As CX Dive put it this month, BPOs make their gross profit on the margin between labor cost in their regions and what they charge per agent. In that version you rent seats and the vendor keeps the spread.

The version that works now looks different. You hire a dedicated, AI-enabled global team that does the work and the review. Agents take the first pass. Global specialists check it against your acceptance test, fix what's wrong and escalate real exceptions. A named owner close to the customer decides what "accepted" means and owns the number.

What you're buying is skilled people at a cost that lets you add review and exception capacity you could never staff at headquarters cost. And your senior people get their time back for the customer. Set it up well and work produced during your day gets checked before your customer's next morning.

One caution. The combination doesn't pay off by itself. In Deloitte's 2024 Global Outsourcing Survey of more than 500 executives (consulting survey), only 26% said AI used by their vendors had reduced the cost of vendor services. Less than half were even asking vendors for lower prices to reflect AI productivity gains. The savings go to whoever measures and contracts for them.

Ask who keeps the savings

If you buy global capacity from a partner, the contract decides who wins. Victor Manzanera, a BPO sector expert, told CX Dive: "If it still pays for hours or seats, nobody in it is paid to save you money."

Before you sign or renew, get answers to these:

• -> When the work gets faster, who keeps the gain?
• -> Who pays for rework on mistakes the AI made?
• -> Who on your side accepts the work, by name?
• -> Will the partner report cost per accepted job, or only hours and seats?

The same test applies to your own team. If a manager's title rises with headcount, you're paying them to hire. Grade them on accepted jobs and cost per accepted job instead.

The worksheet

Run it weekly for one workflow. Use loaded cost (salary, benefits, tools, management time) for every person.

Workflow: ____________________
Accepted job means: ____________________
Accepting owner: ____________________
Week of: ____________________

A. AI and agent spend for this workflow (models, tools, platforms): $______
B. Team cost to do the work, excluding review (hours x loaded hourly cost): $______
C. Review time (hours x loaded hourly cost of each reviewer): $______
D. Rework and retries (fix hours x loaded cost, plus AI spend on reruns): $______
E. Owner escalations (owner and exec hours on exceptions x their loaded cost): $______
F. Total cost (A + B + C + D + E): $______
G. Jobs submitted: ______
H. Jobs accepted: ______
I. Jobs accepted on the first pass: ______

Cost per accepted job = F ÷ H
First-pass acceptance = I ÷ G

Keep one-time setup costs (integrations, prompts, training, hiring) on a separate line. You want to see the recurring number clean, then decide how fast the setup pays back.

Two numbers tell you most of what you need. Cost per accepted job tells you if the model is cheaper. First-pass acceptance tells you if it's getting better or just getting busier.

How to run it in 30 days

1. Pick one workflow with steady volume and a clear yes or no from the customer. Support tickets, release QA, month-end close or SDR research all work.
2. Write the acceptance test. One sentence plus a short checklist. If you can't write it, you can't measure the job.
3. Name one accepting owner. Close to the customer. They own the number and the budget.
4. Measure a two-week baseline with the worksheet before you change anything.
5. Redesign the work. Agents draft. Your global team checks against the acceptance test, fixes and escalates. The owner decides exceptions and changes the test when the customer's needs change.
6. Log every retry and every send-back. Have the agents log attempts and outcomes too. Don't take an agent's word that something is done.
7. Compare after four weeks. If cost per accepted job fell and first-pass acceptance held or rose, add volume. If it didn't, fix the acceptance test or the review step first.

Don't cut headcount on day one. Gartner's advice is to treat reductions as an outcome of proven, sustained AI performance, and the rehiring numbers show why.

Tradeoffs and when this doesn't apply

Some work has no clean acceptance signal. Strategy, brand and some product decisions fall here. Use the owner's sign-off as the proxy and admit it's softer.

Low volume kills the math. If a workflow runs 20 jobs a month, setup cost swamps everything. Start where volume is real.

A global team adds work of its own. Onboarding, overlap hours, security and access rules. Count that time in B and the setup line. Give agents and reviewers only the access the job needs.

The lowest hourly rate can still produce the highest cost per accepted job. A reviewer who misses errors pushes the cost into escalations and into your customer's inbox. That's why the worksheet divides by accepted jobs.

The obvious objection: won't better models make review unnecessary? For some tasks, eventually, yes. Right now the data says human approval is the norm, even at the biggest companies. When your first-pass acceptance gets high enough, review less. The worksheet tells you when you've earned it.

A useful next step

Pick one workflow this week. Write the acceptance test. Name the owner. Fill in the worksheet with last week's numbers. If you can't fill in line H, you don't know what your AI is worth yet.

Disclosure: I'm co-founder and Executive Chairman of Outsorcy, which sponsors this newsletter. Outsorcy places full-time, AI-enabled professionals from Kosovo and Eastern Europe on the teams of mid-market technology companies. Run this worksheet on any model you're weighing, ours included.

Sources

1. Too many scoreboards count speed and headcount cut; too few ask whether outcomes got better: Greg Moran on X, Nov 8 2026. https://x.com/GregMoranX/status/2119422352639283265
2. Gartner survey of 275 service and support leaders (July to August 2026): 21% reduced human agent headcount because of AI; 7% rehired after earlier cuts; 11% plan to rehire within 18 months; "cut fast, rehire later" trap; advice to treat reductions as an outcome of proven AI performance. Gartner press release, Nov 5 2026 (analyst survey). https://www.gartner.com/en/newsroom/press-releases/2026-11-05-gartner-survey-finds-21-percent-of-service-and-support-leaders-have-reduced-human-agent-headcount-because-of-ai
3. Miro survey of 9,500+ leaders and workers: 83% improved individual productivity; 56% improved output quality; 59% more AI-generated output to review; 55% more low-quality AI output requiring review and rework. Miro press release, Nov 3 2026. VENDOR DATA, self-reported perceptions. https://miro.com/newsroom/individual-ai-productivity-soars-but-organizational-impact-stalls/
4. Daily AI spend rose from about $35 to $84 after moving to cheaper models; call volume outgrew price per call; AI fact-checker flagged 18 of 30 accurate notes; "we paid twice for worse copy." Jason Lemkin on X, Nov 7 2026. Operator self-report, not an audit. https://x.com/jasonlk/status/2119140412212310461
5. Jason Lemkin bio: founder of SaaStr (SaaS founder community and conference); CEO and co-founder of EchoSign through its acquisition by Adobe Systems; later VP, Web Services at Adobe. SaaStr company page, read Nov 9 2026. https://www.saastr.com/company-back/
6. TCS Enterprise AI Study, 1,333 senior leaders at organizations with $500M to $20B+ revenue: only 10% let AI agents execute decisions without human approval across multiple functions. TCS press release, Nov 4 2026. VENDOR DATA, enterprise sample. https://www.tcs.com/who-we-are/newsroom/press-release/productivity-top-benefit-tcs-enterprise-ai-study
7. Brian Armstrong's agent setup (lead agents, sub-agents, one prioritized review queue); quotes "Every change still goes through the same human review and security checks" and "The goal is fewer corrections per change"; 14,584 bookmarks at check on Nov 9 2026. Brian Armstrong on X, Nov 7 2026. Founder self-report. https://x.com/brian_armstrong/status/2119112609223827547
8. BPOs make gross profit on the margin between labor cost in their regions and what they charge per agent; most BPO CX revenue still per agent or per seat; Victor Manzanera quote "If it still pays for hours or seats, nobody in it is paid to save you money." CX Dive, Nov 5 2026. https://www.customerexperiencedive.com/news/taskus-ceo-blending-ai-human-talent-working-cx/873945/
9. Deloitte 2024 Global Outsourcing Survey, 500+ executives: 26% report AI used by vendors reduced the cost of vendor services; less than half seek vendor price reductions for AI productivity gains. Consulting survey, published 2024. https://www.deloitte.com/global/en/issues/work/global-outsourcing-survey.html (full report: https://www.deloitte.com/content/dam/assets-shared/docs/services/consulting/2025/global-outsourcing-survey-2024.pdf, p. 7)
10. Outsorcy facts in the disclosure (full-time professionals, AI-enabled teams, Kosovo and Eastern Europe, mid-market technology companies; Greg Moran, co-founder): outsorcy.com, read Nov 9 2026. https://www.outsorcy.com/