AI Credits in Project Management Pricing: The Third Wave of Surprise Billing

Twelve months ago, AI in your project management tool was free, or quietly bundled into the plan you were already on. Six months ago, it became a separate per-seat surcharge — the wave we wrote about in The Hidden AI Subscription Stack. In Q2 2026, the model has shifted again. The new shape of AI credits project management pricing isn't a per-seat add-on at all. It's a meter. You buy a bag of credits, your workspace burns them as agents run, the bag empties, and the bill is whatever your team happened to do that month.

On May 4, 2026 — twelve days ago at time of writing — Notion's Custom Agents flipped from free trial to $10 per 1,000 credits, with one agent run consuming "30 to 60 credits depending on tools, reads, and writes." Asana's AI Studio meters credits in the background of the Starter plan, with automations that silently stop firing mid-month when the pool drains. Atlassian's Rovo charges 10 credits per chat or agent request and 100 per Deep Research call, with consumption pricing for overages already on deck. Three different vendors, the same playbook, all within one quarter. For a small team, this is structurally worse than the per-seat AI surcharge it's replacing, and it's worth understanding why before the next renewal conversation.

AI credits project management pricing comparison across Notion Custom Agents, Asana AI Studio, and Atlassian Rovo showing per-credit costs and unpredictability — The three vendors that moved to credit-metered AI inside one quarter — and the math nobody can forecast

Three Vendors, One Quarter, the Same Bet

The evidence isn't anecdotal. Inside a single quarter, the three largest credit-metered shifts in PM tooling all landed.

Notion Custom Agents went live with a price on 2026-05-04: $10 per 1,000 credits, billed alongside your Notion subscription. Credits are pooled across the entire workspace, reset monthly, and do not roll over. Each agent run burns 30 to 60 credits — meaning a $10 top-up gets you somewhere between 17 and 33 agent executions before you're out. Custom Agents are restricted to Business and Enterprise plans, so this credit meter sits on top of an $18/seat workspace fee you're already paying.

Asana AI Studio ships with 50,000 monthly credits at Starter as the entry-level allowance, then sells AI Studio Plus as a 100K-credit add-on per month and AI Studio Pro at 5M credits per quarter via sales-led pricing. The lower tiers don't publish dollar-per-credit math at all — you discover the rate when your automations stop firing. (Asana help)

Atlassian Rovo charges 10 credits per chat or agent request and 100 per Deep Research call, with monthly allowances ranging from 25 credits per user on Standard up to 1,500 on the Teamwork Collection Enterprise tier. Atlassian is currently letting overage requests through without billing them, but has confirmed consumption pricing is coming with a 90-day heads-up. The trial period is the only thing standing between today's invoice and the metered one.

Three vendors, three different unit names ("credits"), three different conversion rates, one shared shape: the bill is no longer a function of how many people you have. It's a function of how much your AI ran.

Why This Is Structurally Worse Than Per-Seat AI

Per-seat AI was annoying. Credit-metering is destabilizing. Here's the precise difference.

1. Bill volatility kills procurement approval

A five-person team can forecast a per-seat AI bill exactly: headcount times the published rate. They can put that number in a budget, send it to a finance lead, get approval, and move on. A credit-metered bill is a probability distribution. The same team running the same workflows might spend $40 one month and $180 the next, because someone built a Deep Research agent and let it loose on a quarterly planning doc.

This isn't hypothetical. 78% of IT leaders reported unexpected charges from consumption-based or AI pricing models in 2026, and 90% of CIOs cite cost forecasting as their top challenge in AI deployment. The most-cited 2026 cautionary tale is Uber: the company burned through its entire annual AI budget four months in after Claude Code adoption nearly doubled across the workforce. If a company with Uber's procurement infrastructure can't forecast credit consumption, a five-person team won't either.

For small teams, the practical effect is darker than the headline number. Procurement at a five-person company is one founder approving a Stripe charge. That founder can absorb a $40-vs-$80 swing. They can't absorb a $40-vs-$400 swing without learning the lesson that AI features are off-limits — which is exactly the opposite of what the vendor wanted to sell them.

2. Silent failure beats loud failure

Per-seat AI fails loudly. If you stop paying for ClickUp Brain, ClickUp tells you in three places that Brain is off. Credit-metered AI fails silently. You don't get an error when your agent stops running mid-month — the agent just doesn't run. The weekly digest you wired up doesn't get sent. The compliance check you trusted doesn't fire. You discover the credit pool drained when something important didn't happen and nobody noticed.

This is the credit model's deepest design flaw for small teams. A 5-to-15-person team automates AI workflows precisely because nobody has time to babysit them. If "did the credit pool refill yet?" is now a thing somebody has to monitor, the automation has paid back its first month of cost in cognitive overhead before you've even seen the second month's invoice.

3. The "free trial → suddenly $X" pattern is the canonical case

Notion's May 4 flip is the textbook example, but every vendor in this category is running the same playbook. Step one: ship a generous free trial of the AI feature, get teams hooked, let them wire automations into the day-to-day. Step two: announce a pricing model that converts those automations into a metered cost line, with 30 to 90 days' warning. Step three: collect the bill from teams that don't have time to rip the automations out before the meter starts.

We're early enough in the cycle that you can still see the seams. Notion Custom Agents was free for the Business/Enterprise customer until the day it wasn't. Atlassian Rovo is letting overages through for free today and has already confirmed metered overages are coming. The next vendor that ships credit-metered AI as a "free preview" is borrowing the same playbook, and your renewal date will tell you how it ends.

The Numbers Behind the Trend

Two large studies in 2026 line up the same picture. Zylo's SaaS Management Index found 78% of IT leaders experienced unexpected charges from consumption or AI pricing in the past year. McKinsey's 2026 Software Pricing Report found 62% of SaaS platforms introduced AI-premium tiers and buyers reported budgeting 25–35% higher to add AI to existing stacks. (Zylo, 2026)

For a small team, the McKinsey number isn't the headline — it's the floor. A 5-person team running ClickUp Unlimited at $35/month plus ClickUp Brain at $9/seat plus Notion Business at $18/seat plus a Notion credit top-up plus Atlassian Rovo allowance overflow isn't paying 25% more for AI. It's paying multiples more, on bills that arrive in different envelopes.

The Two Ways Out

The exits haven't changed since we wrote about them in The Hidden AI Subscription Stack, but credit-metering makes them sharper.

Way 1: Flat-rate tools where AI is bundled or excluded — never metered

Flat-rate project management tools charge a fixed amount for the whole team. There's no per-seat surface for an AI surcharge to attach to, and no credit meter to drain mid-month. Some flat-rate tools include AI features in the base price; others leave AI out entirely and assume you'll bring your own. Either way, the bill is the same number every month.

The honest tradeoff: vendor-built, deeply-integrated AI workflows (Notion Custom Agents, Asana AI Teammates, ClickUp Super Agents) are not what flat-rate tools usually compete on. If your team has decided proprietary in-product AI is worth a 3x bill and the credit volatility, this exit isn't for you. If your team's AI work happens primarily in Claude or ChatGPT and you mostly want your task tool to stay out of the way, flat-rate is the cleanest exit on the table.

Way 2: MCP — your existing AI subscription reaches into your tools

The second exit is the one that breaks the credit-metering math entirely. Model Context Protocol (MCP) is an open standard that lets the AI assistant you already pay for — Claude Desktop, Cursor, ChatGPT desktop, custom agents — read and write to your tools without the tool vendor having to ship its own AI product. If your task tool exposes an MCP server, your existing AI subscription does the work. The credit meter doesn't exist because the vendor isn't running inference for you.

The arithmetic at the small-team level: Claude Pro is $20/seat/month, covers coding, email triage, research, and PM workflows, and runs against every MCP-enabled tool you adopt. Stacking ClickUp Brain at $9/seat plus Notion credits plus Rovo allowance for the same capabilities scoped one tool at a time is paying multiple times for AI you already have. The MCP route makes your existing AI subscription portable. You pay once. The AI works everywhere.

Heimin's bet is route 2. The entire task system and CRM are exposed as an MCP server, and your AI agent — whichever one you already pay for — can manage tasks, comments, projects, and customer notes through your existing subscription. There's no "Heimin Brain" on the roadmap, no credit meter, no agent run quota. We're explicit about it because we think small teams will, eventually, all do this math.

Practical Takeaways

Map your tools' AI billing model, not just the headline price. A $7/seat tool with a credit-metered AI add-on is not a $7/seat tool. Write down: base price, AI surcharge, AND the credit conversion rate and monthly allowance. The third number is where the volatility hides.
Audit which automations would silently fail if credits ran out. Walk through every AI rule, agent, or automation you've wired up. For each one, ask: "If this stopped firing mid-month, would I notice?" The ones where the answer is "no" are the ones credit-metering will hurt most.
Set a tripwire. If your tool exposes a credit usage dashboard, set a calendar reminder to check it on day 15 of each billing cycle. If you've burned more than half the pool, decide now whether to top up, throttle, or kill the automation.
Run the duplication test. Pick one credit-metered AI feature you pay for today. Ask whether your team's existing AI subscription (Claude, ChatGPT, Copilot) could do the same job over MCP. For most "summarize this week's tasks" or "draft a status update" use cases, the answer is yes — and you're paying twice.
Treat the free trial as a tripwire, not a feature. Any AI feature labeled "free preview" or "included for now" is a credit meter that hasn't started yet. Build automations on top of free trials only if you can rip them out within 30 days of the pricing announcement.

The Heimin Take

We built Heimin's MCP integration before we built any in-product AI features. We took that order deliberately. Our 5-to-20-person customers were already paying for Claude or ChatGPT. Building "Heimin Custom Agents" at $10 per 1,000 credits would have been a clean revenue line for us — and would have meant our customers paying twice for AI capabilities they already had, on a bill they couldn't forecast.

For a small team running the math at renewal, the question isn't whether AI in PM tools is valuable. It often is. The question is whether the AI is being priced in a way you can actually budget around. In 2026, "credit-metered" is the honest answer to "no" for most small teams — and the two exits above are the honest answer to "what do I do about it."