Notion no longer sells AI as “any plan + add-on.” Full access sits on Business / Enterprise. Since May 2026, the interesting part—Custom Agents that schedule, hit Slack / Mail / MCP, and run multi-step work—meters through workspace-shared Notion credits.

Notion’s help center and product page are blunt: monthly credits cost $10 per 1,000; annual credits cost $13 per 1,000 (you pay for in-year flexibility, not a discount). More reading, more steps, more tool calls, stronger models → more burn per run. When the pool hits empty, agents pause; admins get notices around 80% and 100%. Public materials also draw another line: Notion Agent, meeting notes, enterprise search, and other on-demand AI stay included in Business / Enterprise seats—the metered add-on mainly hits Custom Agents that run on their own. The teardown is how a gated seat + metered credits stack into “I thought this was flat”—versus something like Claude’s flat / Cowork-style subscription story.

The easy procurement miss: hearing “Business includes AI” as “agents run free.” Inclusion is the gate and the feature set; Custom Agents burn separate fuel.

Mechanism: shared pool, no rollover, complexity-priced

Official example ranges (approximate; trust your dashboard):

Example agent	Cost per run (approx.)	Runs per 1,000 credits ($10)
Q&A in Slack	~$0.03–$0.11	~90–333
Task routing	~$0.05–$0.15	~65–190
Status update	~$0.08–$0.18	~57–133
Mail triage	~$0.04–$0.10	~100–250
Daily brief	~$0.10–$0.30	~33–100

Notion’s own multiply-out: a status-update-style agent at 60 runs/month lands about $4.80–$10.80—still “small miscellaneous” money. Flip to a weekday daily brief (~22 runs) at ~$0.20 mid → ~$4.4 / month; add mail triage (~20/day × $0.07 ≈ $42) and task routing (~30/day × $0.10 ≈ $90) and a three-pack is easily $130+ in credits—on top of seats.

Same goal, a complex router (multi-DB search, sub-page, Slack notify) costs more than “read one message → create one task.” The help center contrasts simple vs complex task routing explicitly: more DBs, more fields, more sub-pages, more Slack fan-out → more credits at the same run count. Credits are workspace-shared—every agent draws one pool; mid-cycle increases apply immediately, decreases next period. Downgrade to Free / Plus turns agents off (not deleted). App Store / Play subscribers cannot buy credits in-product—they must move to web Business/Enterprise.

Also easy to miss: Basic Autofill does not use credits; Custom Agent Autofill does. Teams that assume “auto-fill columns are free” discover metering only after picking the Custom path. Third-party guides (e.g. Matthias Frank) add an ops detail: when credits run out, scheduled runs are skipped—they do not catch up later. A pause on the 19th is not only a feature freeze; it is a hole in the production calendar. The admin guide adds another gate: per-agent credit caps (Enterprise can set defaults) so one experimental agent cannot drain the whole workspace.

AI Vault’s 2026 review adds the procurement reality: the standalone AI add-on is gone; Business at about $20 / member / month is the AI gate. Free / Plus keep a one-time trial batch, then lock. Custom Agents had a free window through May 3, then credits from May 4. The review roughly maps Notion’s examples to ~17–33¢ per run; a weekday morning digest can stay in single-digit dollars, but digest + triage + weekly report stacks into tens of dollars of credits—still on top of seats.

Spreadsheet-shaped scenarios:

Scenario	Seats	Credits estimate	~All-in / mo	Contrast
Solo upgrades for “better writing”	1 × $20	First 1,000 ($10); narrow Q&A often underspends	from $30	Flat Claude Pro ~$20 often covers general writing
Five people, three scheduled jobs	5 × $20 = $100	~15 runs/day × $0.15 → ~$67.5 (buy ~7,000 ≈ $70)	~$170	Five Claude Pro seats ≈$100—but no unattended Notion-wiki automation
Twenty people, agents as prod ops	20 × $20 = $400	~80 runs/day × $0.12 → ~$288 (~30k credits ≈ $300)	~$700	Five Claude Max seats ($100–$200 each) may or may not be cheaper—but the budget shape is capped

The solo line is the quiet trap: Business ($20) + 1,000 credits ($10) starts at $30 before premium models and multi-tool chains—easy to outspend flat Claude + free/Plus Notion. The five-person line flips: ~$170 often still beats five heavy flat seats—if agents truly run workspace work instead of full-wiki scans.

Monthly $10/1,000 vs annual $13/1,000 is a real choice. Help center: annual is not cheaper—you pay for draw-anytime flexibility across the subscription year; monthly fits steady load and resets. Self-serve: monthly Notion plans buy monthly credits; annual plans buy annual credits; account-managed orgs can add on-demand. Draw order across pools is roughly free → monthly → annual. Hoarding a year of credits is not a savings strategy—neither type rolls over; annual only stretches the evaporation window to renewal.

Extreme burn anecdotes in the review (one agent ~3,000 credits/month ≈ $30; a team at six-figure credits) are unverified invoices—but the shape matches Notion’s own drivers: read-everything, many connectors, many steps, expensive models. Notion’s own advice—narrow scope, measure 1–2 billing cycles—is the only honest onboarding for a metered product.

Business model: surprise pause vs flat Claude / Cowork

Surprises rarely mean a silent card charge—Notion warns at 80% and pauses at empty, sold in blocks. The real shock is agents stop on the 19th (schedules skipped), or three scheduled jobs turn a mental “$10” into tens of dollars of credits—on top of Business seats. No rollover punishes hoarding: oversized blocks evaporate; undersized blocks halt mid-month. Finance wants predictable; the product offers pauseable—neither equals flat.

Contrast a common alternative: Claude’s flat / Cowork narrative—public bands roughly Pro ~$20 / month, Max ~$100 / $200 (usage multiples, not a separate Cowork SKU), Team per-seat caps. Spreadsheet-friendly: hit the ceiling and you stop for the day; you do not silently skip Tuesday’s scheduled run. Notion Custom Agents move “will this runaway?” from model IQ to workflow surface area. AI Vault’s product point: workspace retrieval and scheduled ops are Notion’s edge; for general writing/reasoning, a flat assistant is often cleaner—and many buyers stack both.

Procurement sentences that hold:

• Human-in-the-loop completion → flat Claude / Cowork: cost follows the person; stop using, stop burning.
• Unattended runs that must read the Notion wiki → Business gate + credits: you are buying calendar automation, not another chat window.
• Both → usually stack; don’t burn credits on general writing, and don’t expect Pro chat to own Monday’s status digest.

Build order moves the bill more than the prompt. Notion lists “how much you read” first among burn drivers: an agent that scans the whole workspace pays for irrelevant pages every run; pin it to one database or teamspace and the same output often costs an order of magnitude less. Tool calls rhyme—one agent chaining Notion → Slack → Linear → Calendar makes four round trips; three narrow agents are usually cheaper and easier to debug. Model tier is another lever: a mechanical daily digest does not need the strongest model; pay for premium reasoning only when projects conflict. Third-party notes also flag Effort (reasoning intensity) as a per-run token dial—set it deliberately on trigger-fired agents. That is not a tip sheet—it is the metered product rewriting workflow design as a cost function.

A contrast with Fin-style outcomes helps: Fin bills “successful results” (watch the definition); Notion bills “how much work ran” (watch scope and steps). Neither is flat, but runaway modes differ—assumed resolutions inflate one bill; over-wide, over-frequent agents drain the other.

What the Business gate actually means

The gate is not only “$20 more.” It decides who is allowed into the metering game:

1. Free / Plus: Custom Agents are off the table; downgrade turns agents off but keeps them around—you lose runtime, not the config archive. The “try agents free, then maybe upgrade” window mostly closed when credits went live in May 2026.
2. Business seats are sunk: full AI (on-demand Notion Agent, meeting notes, etc.) rides that ticket; credits are ticket two. If procurement approved “AI plan” but not “automation usage,” month two stalls in the add-on workflow.
3. Mobile-store subscribers: cannot buy credits in-product; must move to notion.com Business/Enterprise—real friction for solo builders, not a footnote.
4. Shared-pool politics: credits ignore who created the agent; a wide-scanning sales bot can starve engineering’s morning digest. Enterprise per-agent caps push spend control down to agent granularity.

So half of “should we buy credits?” is really “should we stay on Business?” If the main value is docs + occasional AI Q&A, included on-demand AI may be enough—don’t keep buying fuel for demo agents. If the value is “wiki state becomes Slack/mail actions every day without a human,” credits are production cost, not optional.

Where intuition fails

Intuition 1: $20 Business = unlimited agents.
$20 buys the gate and workspace AI; Custom Agents are a second bill. Notion’s product page is explicit: seat price stays; other AI features remain included—metering hits proactive Custom Agents.

Intuition 2: metered always beats flat.
Often true for low-frequency, narrow scope; high-frequency, multi-tool, premium-model flows can make flat Claude calmer. Comparing “five Business + $70 credits” to “five Pro” still requires asking whether the savings bought unattended wiki work.

Intuition 3: buy a fat credit buffer so nothing pauses.
No rollover turns safety stock into waste; size to measured use and top up. Skipped schedules on empty do not get backfilled—hoarding cannot fix an over-wide agent. Annual $13/1,000 also does not roll over; it only stretches the flexible pool to renewal.

Who should do what · who should buy credits

Role	Math to run	Don’t mythologize
Teams already on Business	Run one agent for a full billing cycle before extrapolating; narrow data scope, fewer tools, prefer Auto; name a credits owner; cap expensive agents	Ship five schedules in week one and read the bill in week five
Plus / solo users	Whether Business seats + credits buy “answers from our wiki, unattended”; compare all-in vs flat Claude (Pro ~$20 or Max); App Store subscribers price the migration	Treat $20 as “another ChatGPT”; keep 5,000 credits for occasional hand-built agents
Finance / procurement	Seats are the gate floor; credits are the slope; monthly for steady, annual for spikes; who answers the 80% alert	Write credits as “unlimited AI flat”
Agent-product builders	Pause-over-overdraft is a trust feature; no rollover pushes honest sizing; the Business gate decides whether you serve teams or solos; skipped schedules are ops incidents	Assume metered always beats flat—high-frequency multi-step can invert

Buy credits when: triggers are stable (schedule/Slack/DB), scope can be pinned, per-run cost is repeatable from the dashboard, and pause cost exceeds top-up friction.
Don’t—or buy the smallest block—when: you are still prompt-debugging, default visibility is the whole wiki, nobody owns failed runs, or the real job is general writing (leave that on a flat assistant).

Credits turn Custom Agents from a demo into an auditable ops line. Only what fits the budget table stays on the production calendar.

Sources

• Buy and track Notion credits for Custom Agents (Notion Help)
• What are Notion credits? (Notion Help)
• Notion AI Review 2026 (AI Vault)