How much should you spend on AI tools per engineer?

Short answer. Expect $20–60 per engineer per month if you buy one seat — a paid license for one person — and nothing else. Expect $150–450 if your team also runs agent tools, the kind that work through a job on their own and bill by the token. A token is the unit AI vendors bill for, roughly a few characters of text. Those two bands are estimated, not counted. We built them from published list prices. The version by company stage is further down.

You cannot just look this up, and the public benchmarks that do exist describe other companies, not yours. DX, in June 2026, puts the total per engineer at $200–600 per month — seats plus tokens, on teams that mix inline tools with agent tools. Read the shape of even that one figure: a 3x spread from the bottom of the band to the top, inside a single source.

So the honest answer is not a number, it is a range that belongs to a population you may not sit in. The useful move is to understand why any per-engineer benchmark is this wide, then place yourself against your own books instead of someone else's sample. Here is how.

Start with the shape of the spread

Take the range seriously, not just its midpoint. DX's $200–600 is a 3x spread on its own, and it only covers teams already running agent tools — the light, inline-only shops sit below it, and an agent-heavy tail sits above. A percentile is where you sit if you lined every company up in order; any figure that collapses the whole spread into one "average AI spend per engineer" throws away the percentile you actually needed.

The spread is not noise. A few firms running heavy agent or training work sit an order of magnitude above the rest, and any average drags toward them — somewhere no normal company lives. That is the direction a single headline number misleads you: high, toward a tail you are almost certainly not in.

So the first rule is a rule about words. Stop asking for "average AI spend per engineer." The average is a fact about outliers. A range, and where you sit inside it, are facts about you.

Five reasons the per-engineer numbers scatter

1. They divide by different things. One benchmark divides by every developer on the payroll; another, like DX, describes cost per engineer only on teams that have already picked up agent tools. Say 60% of your engineers hold an AI seat. Your per-developer number and your per-holder number are then 1.7x apart, before anything else moves. Most published benchmarks never say which one they used.

2. Seats only, or seats plus tokens. GitHub Copilot Business is $19 per user. Cursor Business is $40 per user. Claude Code Team Premium is $100 per seat. A seat-only benchmark stops there. Agent tools then burn tokens on top of the seat. Anthropic lists Claude Opus 4.8 at $5 in / $25 out per million tokens, and Claude Sonnet 5 at $2 / $10 (read July 2026). A seat-only figure cannot see that line at all. It is the same gap that makes pay-per-use pricing so easy to under-forecast.

3. Who answers the survey. Firms that answer an AI spend survey tend to have AI spend worth reporting. They also tend to own the tools to report it well. No published index is a random sample of the economy.

4. Tool mix. Inline code completion and agent work cost money in different ways. Copilot-style completion is close to a fixed seat cost. Claude Code runs about $13 per developer per active day. Call it $150–250 a month for a daily user. A team that is 90% inline and a team that is 90% agent can be tenfold apart at the same headcount.

5. Where the tail sits. The top of any published range includes firms whose "AI spend" quietly bundles in model hosting or training. That has nothing to do with coding tools. If a benchmark's top decile counts it and yours does not, you are not comparing the same thing.

The mean of an AI spend spread is a fact about outliers, not a fact about you.

Four rules for placing yourself

Use a range and a position in it, never the mean. Say "our per-holder spend sits near the top of the observed band." Do not say "we are below average." The first can be checked against your own books. The second is noise.

Compare seats to seats. First split your own spend into seats and usage. Then hold each line against a benchmark that counts the same line.

Keep inline tools and agent tools apart. Report them as two rows. Their cost per unit differs, their take-up differs, and the renewal argument differs.

Say out loud what you divided by. Per engineer on payroll, per engineer with a seat, or per engineer who used a seat in the last 30 days. Those are three different numbers. The third is usually the one that matters. How many seats get used moves your per-engineer cost more than pricing does — see what good AI seat utilization actually looks like.

A read by stage — estimated, not counted

The table below is estimated. We built a likely tool stack for each stage out of published list prices. It is not survey data. No company was counted to produce it. Treat it as a starting guess you replace with your own invoices.

| Stage | Typical stack | Estimated seat cost | Estimated usage cost | Estimated total / engineer / mo | |---|---|---|---|---| | Seed–Series B | One inline assistant. Copilot Business $19 or Cursor Pro $20 | $19–40 | $0–20 | $20–60 | | Series C–growth | Inline plus one agent tool, partial overlap. Cursor Business $40 (Vendr-observed negotiated $30–45), Claude Code Team Premium $100/seat | $40–140 | $20–120 | $60–260 | | Enterprise | Copilot Enterprise $39 + $21 GHEC = $60 effective, agent seats alongside, plus direct API or Bedrock usage | $60–200 | $50–250 | $110–450 |

Two things to notice. The seed band sits well below DX's range, and the enterprise band overlaps the bottom of it. That is what you would expect: DX's population is teams already running agent tools, so a seed shop on one inline seat should land beneath them. Also, the usage column holds the real spread. It is the column your finance system probably cannot break out by team today.

Cross-check it against budget share

A per-engineer number is easy to game by moving headcount around. Pair it with a share-of-budget check. Cledara's live data hub put AI at 16.9% of total software budget as of July 2026. Say your per-engineer figure looks normal but AI eats 35% of software spend. One of the two numbers is lying to you. The share-of-engineering-budget view is the other half of this check.

A benchmark tells you if you are odd, not if you are right

This is the part that gets skipped. A percentile tells you where you sit against a sample. It says nothing about whether the money is working.

KPMG's Global AI Pulse, Q2 2026 asked 2,145 C-suite leaders across 20 countries. Only 7% report established ROI on AI. Among leaders who can see their costs clearly, 15%. Among those who cannot, 3%. Clear cost data raises the odds about five times. And 42% can only partly see what they spend.

So the number that counts is spend against output, not spend against a peer group. DX's AI Efficiency Plateau study covered 400+ companies from November 2024 to February 2026. The median gain in pull requests shipped was 7.76%. The bottom 10% sat at −3%. The top 10% sat at +44%. That spread is wider than the spend spread. Two firms spending the same amount per developer can sit at either end of it.

Throughput — how much work gets finished — is not the whole answer. Read it next to the cost of going faster: rework, which is work that has to be redone, plus bugs per pull request, review time and incidents. A speed gain bought with four times the review load is not a gain. Forcing that pairing is the job of an AI cost report, one page that shows what your AI tools cost and what you get back.

The honest caveats

The one benchmark here is not a random sample. DX's range is what its own people see across client teams, not a published spread, and it leans toward firms organised enough to count. The stage table is estimated from list prices. It is wrong for any company with real negotiated discounts. Vendr's Cursor marketplace shows 15–25% off for an annual commit, and 25–40% when you pull more than one lever. That is enough to move a whole row. Nothing here counts AI hosting outside developer tools. And a benchmark drawn from mid-2026 prices will age fast. Prices per token keep falling while tokens used per task keep rising. Work it out again rather than quote it again.

Find your recoverable AI spend

Your percentile only gets interesting once you know how much of that spend sits on seats nobody opened last month. Spendassay reads seats and usage across every AI vendor at once. It reports your real cost per engineer against the people who actually used the tools — see how the proof levels work.

`Start free` → /login?src=blog_ai-spend-per-engineer-benchmark