Our founder spent a Saturday morning building a graphics workflow on Claude Design at the $100/month plan, made it about ten outputs in, then got locked out for a full week. The reset clock said five hours. The reality was 165. The temptation to upgrade to the $200 tier was immediate. That experience is not unique to one tool, and it is not an accident. Credit-based AI pricing is becoming the most subtle vendor lock-in pattern of the current cycle, and construction firms quietly building AI workflows on top of it should be paying attention.
The story starts on a Saturday at noon. Owen, our founder, was experimenting with Claude Design, turning podcast quotes into shareable graphics for the team. The plan was to build a repeatable, idiot-proof process so anyone could produce visuals from raw content. After about ten outputs, the screen flashed up the message every AI power user is starting to recognise. Credits exhausted. Next reset at 5:05.
5:05 looked fine. Five hours away. Then he looked again. The reset was 5:05 in the morning, the following Saturday. A full week. The path to keep working was a one-click upgrade from the $100 plan to the $200 plan, conveniently surfaced exactly at the moment the workflow stalled.
That moment is the article. Whether you run a ConTech startup, a contracting firm, or a construction software company embedding AI, you will encounter some version of this. The longer you build on top of credit-based AI, the more expensive switching becomes.
Credit-based pricing is not inherently bad. It is what credit-based pricing does to your operational dependence over time that matters. The lock-in is not the contract. It is the workflow.
The week-long lockout
What actually happens when credits run out mid-workflow
The mechanic is straightforward. You pay a monthly subscription. The subscription comes with a bucket of credits. Credits get consumed at a rate the vendor decides, against a token-counting formula most users do not fully understand. When the bucket is empty, you wait, or you upgrade.
The wait is the part that quietly does the work. A five-hour reset is annoying but tolerable. A 165-hour reset is a different category of problem. By the time the credits return, the workflow you were building has stalled, and your only option to keep momentum is the upgrade button.
Once you have built a workflow on top of a tool, every credit reset becomes a more expensive problem. The first time you hit a wall, it costs you a Saturday. The tenth time, it costs you a publication schedule, a client deliverable, or a tender deadline. The vendor is not doing anything dramatic. They are just letting the cost of your dependence accumulate.
This is not a Claude story
Why every major AI tool is on the same trajectory
The specific tool here is incidental. The pattern is industry-wide. GitHub Copilot announced in late April that all plans are transitioning to fully usage-based billing on June 1, 2026, with monthly allotments measured in AI Credits and overages billed by token consumption. Developer reaction has been pointed: predicting how much work a subscription will do becomes substantially harder, even though headline subscription prices have not moved.
Microsoft 365 deducts AI credits per action. Google AI plans run on tiered credit baselines with paid top-ups. OpenAI uses a hybrid of seat fees plus credit packs. HubSpot’s AI tooling uses pre-committed credit blocks. Vendors have decided credits are the unit of pricing for AI features.
None of this is conspiracy. AI inference is genuinely expensive and genuinely volatile. A single user running an agentic workflow can burn through more compute in a day than a thousand people doing autocomplete. The problem is the secondary effect on customers, especially in industries like construction where digital workflows are still being built and switching costs compound.
If you become too reliant on these tools, then they will just lock you in at some point and force you to continue to upgrade to facilitate your lifestyle.Owen Drury, Bricks & Bytes podcast
The compute crunch makes it worse
Why rate-limiting is happening even before pricing changes hit
There is a second dynamic underneath the pricing one. AI infrastructure capacity is constrained. In late March 2026, Scientific American reported that heavy users of Claude were burning through five-hour usage windows in 20 minutes during peak hours, and Anthropic explicitly warned subscribers that high-demand sessions would exhaust limits faster. OpenAI has begun shutting Sora down at peak times as Codex usage hit four million developers per week.
For end users this means the credit you bought may not always do the work the brochure said it would, because rate-limiting is increasingly happening inside the credit window itself. Vendors are choosing rate limits over price hikes for now, which pushes the friction onto the moment of usage rather than the moment of purchase.
For construction firms experimenting with AI agents for takeoffs, RFI sorting, or document review, this matters. The team that has built a five-step automation on top of an AI tool needs a fallback for the day the credits run out at hour two of a deadline.
Hear Owen’s full story on the latest episode
The credit lockout, the upgrade temptation, and what it signals about AI dependency. The full conversation is on Bricks & Bytes.
Listen to the episodeWhat this means for construction firms
The dependency map most firms have not drawn yet
Construction adoption of AI is further along than most narratives suggest. Recent BuildOps research found 78% of trade contractors are already using AI in some form. That is not pilots. That is in-production usage. The credit-based pricing model is now embedded in tools real construction businesses depend on for daily work.
The question to ask, if you run any of these workflows, is simple. If your AI tool’s credit bucket emptied tomorrow morning at the start of a tender deadline, what would actually happen? In most firms the honest answer is: the work stops, someone scrambles to upgrade, and the budget line for AI tools jumps a tier without anyone noticing. We touched on this in our piece on how construction’s coordination problem is becoming a budget problem.
Three guardrails are worth putting in place now. Set a monthly cap that triggers an alert before the auto-upgrade button appears. Track which workflows touch which tools, so you know what stops if a credit bucket empties. And keep at least one alternative tool warm enough that a forced switch is a half-day inconvenience, not a week-long crisis.
The dependency map at a glance
| Risk | Symptom | Mitigation |
|---|---|---|
| Credit lockout mid-workflow | Team blocked, deadline pressure, upgrade button surfaced | Cap usage with alerts before exhaustion; keep an alternative tool warm |
| Auto-upgrade with no downgrade right | Locked into higher tier even if usage drops | Negotiate bilateral flexibility at renewal; review tier monthly |
| Token-cost opacity | Hard to predict next month’s spend | Track per-workflow token usage; demand consumption dashboards |
| Compute crunch rate-limiting | Credit “available” but performance degraded | Plan workflows around peak hours; build async fallbacks |
| Vendor pricing changes mid-cycle | Credit value drops without notice | Negotiate price locks; ask for change-notice clauses in writing |
Frequently asked questions
It is not lock-in in the strict legal or contractual sense, since most plans are month-to-month. The lock-in is operational. Once your team has built workflows on top of a specific tool, switching means rebuilding those workflows, retraining the team, and absorbing a productivity gap during the transition. Credit pricing accelerates that effect because it makes upgrade pressure feel like a workflow problem rather than a procurement decision. (Source)
The honest answer is that AI inference is expensive and highly variable. One user running a complex agentic workflow can consume more compute in a day than a thousand light users in a month. Credit-based billing aligns vendor cost with revenue. GitHub stated this explicitly in their April 2026 transition announcement: usage will be calculated on token consumption to make Copilot’s economics sustainable. The pricing logic is reasonable. The customer experience is the problem. (Source)
Yes. Scientific American reported in April 2026 that heavy users of Anthropic’s Claude were exhausting five-hour usage windows in around 20 minutes during peak hours, with the company telling subscribers that sessions would burn through limits faster during high-demand periods. OpenAI has been temporarily shutting Sora down to manage capacity. The pattern is industry-wide and reflects a real compute supply constraint. (Source)
The five most useful are: can you set hard caps with alerts before auto-upgrade kicks in? Do unused credits roll over or expire? Can the vendor lock the per-credit price for the contract term? Can you downgrade mid-contract if usage drops? And what happens if the model or feature changes mid-cycle? If a vendor cannot answer those clearly in writing, that is itself a useful piece of information. (Source)
It is a real concern for construction. Recent BuildOps research found 78% of trade contractors are already using AI in some form, much of it priced on credit-based or usage-based models. As workflows for takeoffs, document review, RFI sorting, and proposal generation get built on these tools, the operational dependency on a vendor’s credit policies becomes a real procurement question rather than a hypothetical one. (Source)
Probably not in the short term. PricingSaaS tracked over 1,800 pricing changes across the top 500 SaaS and AI companies in 2025 alone, an average of 3.6 per company. The volatility is structural, not transitional. Buyers should expect more frequent price adjustments, not fewer, until vendors have figured out how to map AI costs to outcomes the customer can verify. (Source)
Related Articles
78% of Trade Contractors Use AI: Adoption Is Way Ahead of What You Think
Construction’s Coordination Problem Is Now a Budget Problem
Why AI in Construction Is Focused on the Wrong Problem
