Runs never guess
Every step executes exactly as written. No run acts on an AI answer, so the same test on the same data gives the same result.
AI in AutoTestX
AI that drafts, heals and explains. People who decide.
AutoTestX uses AI where it saves your team time: writing steps, finding what moved, explaining what broke, and answering your AI assistant. The runs themselves stay deterministic, and every change to a test is approved by a person.
Test case, as your team wrote it
- 1Sign in to the customer portal as a support agent
- 2Open a new service request
- 3Choose the category Facilities
- 4Describe the problem: Heating not working on floor 3
- 5Submit the request
- 6Check the message says the request was created
✦ Draft steps
- Sign in as
{{secret.AGENT_USER}}from line 1 - Click
New requestfrom line 2 - Select Category
Facilitiesfrom line 3 - Type Description
Heating not working on floor 3from line 4 - Click
Submitfrom line 5 - Expect message
Request SR-… createdassumed · check
✦ 6 steps drafted1 assumption to reviewReview and save
Run #2187 · Submit a service requestStaging
- 01Open
Customer portal✓ - 02Sign in as
{{secret.AGENT_USER}}✓ - 03Click
New requestHealed: #btn_new_7f2 → button “Create request” · 96% match · waiting for your approval✦ - 04Select Category
Facilities✓ - 05Type Description
row 3 of requests.xlsx✓ - 06Click
Submit✓ - 07Expect message
Request SR-1042 created✓
7 passed✦ 1 healed · approve fix7 screenshots
› Run the Checkout tests on Staging and tell me if anything broke.
autotestx · run_suite Checkout on Staging · 12 tests
11 passed · 1 failed · 2 min 41 s
autotestx · explain_failure Apply discount code
Apply discount code failed at step 4: it expected “Discount applied”, but the page said “Code expired”.
The button and the form worked, so this is the test data: the promo code has expired. It passed until 29 Sep. Renew it in the staging seed data.
Failure class · Test dataEvidence · step 4 screenshot
Illustrations of the AI step writer, a healed run, and an AI assistant using the AutoTestX MCP server.
Suggestions, not changes
AI drafts, suggests and explains. A person approves anything that changes a test.
Shown its working
Each AI answer comes with the evidence it rests on, beside the deterministic result it never replaces.
01 · AI step writer
From a written test case to draft steps.
Paste a test case the way you would explain it to a colleague, one action or check per line. The step writer drafts the steps, reusing the fields your saved tests already find. Each draft step shows the sentence it came from, and flags anything it had to assume.
- Nothing is saved until someone reviews the draft
- Secret values are never sent: it sees secret names only
- Learns your project’s wording from the tests you keep
02 · AI healing
When the rules can’t prove which element moved.
Self-healing first tries free, deterministic matching. When that finds look-alikes but can’t prove one, for example “Purchase Requisition” renamed to “PR Request”, an AI is asked to choose among the candidates, or to say none of them.
- It sees the step and the candidates’ roles, names, labels, ids and headings
- It never sees the page, a screenshot, a typed value or a secret
- It never writes a locator; the pick is filed as a suggestion for a person
- The same question is asked once a week, however many tests hit it
Tier 2 · AI pickmedium confidence
Step 4 · Click Purchase Requisition
- 1link “PR Request” Tier 1: 58%
- 2link “PR Reports” Tier 1: 54%
- 3button “Print” Tier 1: 31%
The AI compared: “name=PR Request” · “name=Purchase Requisition”. Same menu, same position.
ApprovePick insteadReject
Illustration of a suggestion card
03 · Failure classes and AI diagnosis
Why it failed, before anyone opens the log.
Every failure is classed by rules first, from what the run recorded: the error, failed requests, messages on the page and the test’s history. When the rules can’t decide, AI diagnosis explains the likely cause in plain language, beside the class it never replaces.
Wrong valueLocatorTimeoutServer errorAuthentication / permissionTest dataEnvironmentMaximo business rule
- It sees named details of the failure: the step, the error, failed requests
- It never sees screenshots, recordings or secret values
FailedApply discount codestep 4 of 6
Expected“Discount applied”
Actual“Code expired”
Class · Test datapassed until 29 Sep
✦ AI diagnosis
The button and the form worked; the promo code SAVE10 has expired. Renew it in the staging seed data, or use a code that does not expire.
Illustration of a failed step
04 · AI assistants (MCP)
Ask your coding assistant whether the release still works.
AutoTestX includes an MCP server, the open standard AI assistants use to call other tools. Connect Claude Code, Cursor, VS Code (GitHub Copilot) or Claude Desktop once, with an API token, then ask in plain words. The assistant runs the suite on the AutoTestX server, reads the evidence of each failure, and in an editor can fix the code and run the suite again.
Connect your assistant →Tools your assistant gets
- list_suites
- Find the suites in a project
- run_suite
- Start a suite on an environment and wait for the result
- get_result
- Read a run’s outcome
- explain_failure
- Explain why one test failed, from its evidence
- investigate_failures
- Go through every failure of a run
- release_report
- Summarise a release: regressions, new and known failures
Control
Your key, your switch, your allowance.
Every AI feature runs on one Anthropic key per organization, chosen by an organization admin. Changes are written to the audit log, without the key.
| Choice | AI runs on | Counts against the plan |
|---|---|---|
| Your organization’s own Anthropic key | Your Anthropic account | No. Usage is still listed. |
| The AutoTestX key (default, except on Enterprise) | AutoTestX’s Anthropic account | Yes |
| AI off | Nothing | – |
| Plan | AI generations |
|---|---|
| Free | 100 a month |
| Team | 1,500 a month |
| Business | 5,000 a month |
| Enterprise | Your own key, no AutoTestX limit |
A saved key is stored encrypted and shown only as its last four characters. The step writer, diagnosis and AI healing share one allowance. AI warns at 80% of it, keeps working to 110%, then pauses until the reset or a top-up; building and running tests never pause. One person can make up to 100 AI requests a day. Plans →
FAQ
Questions about AI.
Can we turn AI off?
Yes. An organization admin can turn AI off for every project. AI healing is also off in each project until an Admin or Test Manager turns it on. Everything else in AutoTestX works without AI.
Does AI ever change a test on its own?
No. The step writer’s drafts are saved only when a person saves them. An AI healing pick is filed as a suggestion and never acted on during a run, and self-healing rules never apply one.
What is sent to the AI?
Only what each feature needs: your test case text and known field names (step writer), the candidate elements’ roles, labels and ids (healing), the named details of a failure (diagnosis). Never screenshots, recordings or secret values.
Which model?
Anthropic’s Claude, through Anthropic’s API, on the AutoTestX key or your own. Amazon Bedrock and Google Vertex are planned.
What counts as an AI generation?
One successful AI request. Failed requests are not counted, and on your own key nothing counts against the plan.
Request a session
Bring one workflow. See the AI test it.
In 30 minutes: a live demo of AI drafting, healing and explaining tests on a representative flow, and what your team would need for a pilot.