Operations & Internal Tools

Connect OpenAI and Retool

AI does the reading and matching, and your team checks the results in one internal tool.

Scope this integration Book a discovery call

Connects: OpenAI and Retool
Data flows: 4 automated, two-way where safe
Built with: OpenAI, Retool, SAP
Timeline: Live in 3 to 5 business days from scope
Owned by: You, with accounts, credentials and code in your name.

In short

Connecting OpenAI to Retool puts AI steps inside an internal tool where staff can see and correct the results. SBD built an enquiry platform for an industrial distributor that reads emailed RFQs with AI and matches parts against the SAP item master, with OpenAI and Retool in the stack.

The workflows

What runs between OpenAI and Retool.

Each line is one automated workflow. It shows the event in one tool, the checks and mapping in between, the action in the other tool and who is notified.

RFQ reading and matching
When
A document or email reaches the tool
Automation
- Verify and validate
- Map and normalise
- Upsert by key
Then
Extract fields as structured output and record the model used
And
- Audit log entry
- Alert if it fails or goes quiet
- Review queue if unsure
Purchase order extraction
When
Items are extracted
Automation
- Verify and validate
- Map and normalise
- Upsert by key
Then
Match them against the item master and score the match
And
- Audit log entry
- Alert if it fails or goes quiet
- Review queue if unsure
Invoice coding for review
When
A match is uncertain
Automation
- Verify and validate
- Map and normalise
- Upsert by key
Then
Send it to the review queue, with the source beside it
And
- Audit log entry
- Alert if it fails or goes quiet
- Review queue if unsure
Quote drafting
When
Staff correct a result
Automation
- Verify and validate
- Map and normalise
- Upsert by key
Then
Save the correction and use it to test later changes
And
- Audit log entry
- Alert if it fails or goes quiet
- Review queue if unsure

Every record passes through the same stages

1Receive
Webhook or polling from the source tool, with a signed payload and a replay log.
2Validate
We check required fields, types and the record key. Anything that fails goes to the review queue.
3Map
Fields are mapped to the destination schema, with dropdown values, currencies and dates normalised.
4Upsert
Each write is matched on the record key, so a retry updates the existing record.
5Confirm
The outcome is logged, an alert fires if the count stops moving, and a reconciliation runs each night.

Read this first

Where the simple setup breaks down.

A one-step Zap connects OpenAI and Retool in an afternoon. This is what tends to go wrong once volume grows and edge cases appear.

Free text cannot drive a workflow
A paragraph from a model cannot update a record reliably. Requesting structured output against a schema means bad results are caught instead of reaching the next step.
The model does not know your catalogue
Asked alone, a model will produce a plausible part number. Matching should run against your actual item master, with the model helping to interpret what was asked for.
People need to see why
A tool showing only the answer gives staff nothing to check. Showing the source text beside each result makes review quick, and it is how they come to trust it.
Costs and changes need logging
Without a record of the model, input and result, you cannot explain an error or spot rising cost. Each call gets logged, and corrections become test cases.

What we wire up

The data flows.

01
When
A document or email reaches the tool

Then
Extract fields as structured output and record the model used
02
When
Items are extracted

Then
Match them against the item master and score the match
03
When
A match is uncertain

Then
Send it to the review queue, with the source beside it
04
When
Staff correct a result

Then
Save the correction and use it to test later changes

Used for

RFQ reading and matching
Purchase order extraction
Invoice coding for review
Quote drafting
Structured outputs in Retool
Review queues
Confidence scoring
Item master matching
AI call logging
Evaluation from corrections

Built with

OpenAI
Retool
SAP
Mailgun
Cloudflare

In every build

What every integration build includes.

Under the hood

Upserts keyed on the record
Each write is matched on the source ID, so a retry or a re-fired event updates the existing record.
Retries with backoff
Rate limits and timeouts are retried with increasing delays. If they still fail, they are parked for review.
Idempotent, loop-safe writes
Two-way flows carry a marker so a change written by the sync does not re-trigger the sync.
Reconciliation
A nightly job compares counts and totals on both sides and lists any record that differs.
Audit log
Each run records the input, the mapping result, the action taken and what triggered it. You can search it by record.
Error monitoring
Failed runs raise an alert. An outcome count also alerts when it stops moving, which catches failures that throw no error.
Review queue
Records the sync is not sure about land in a screen a person can read, correct and release.
Scoped credentials
Each connection has its own credential, limited to the objects it needs and stored encrypted in your account.

For the team

Slack or email notifications
Alerts go to a named person on your team, by Slack, email or WhatsApp.
A runbook
One document per integration: what it does, what each alert means, and what to do in the first ten minutes.
Training
A live session with the people who will use the review queue, recorded for whoever joins later.
Handover
Admin access, billing and code sit in accounts in your name from the start, so the build keeps running if we stop working together.
Sixty days of hypercare
The engineers who built it stay on call after go-live, fix what breaks and tune what the team asks for.
A monthly report
One page in plain language covering runs, records synced, exceptions handled and what changed.

Security and your data

Where the data goes, and who holds the keys.

Accounts: Every platform account is created in your name. We work through managed access you can revoke.
Credentials: Stored encrypted in the orchestration layer, scoped per connection and rotated at handover.
Data in transit: All traffic uses TLS, and webhooks are checked by signature before anything is processed.
Data at rest: We store only what the sync needs and keep no copies of your customer data on our machines.
AI steps: Where an integration uses a model, it is through a business API that does not train on your data.
Retention: Run logs are kept 90 days by default, which you can change. Card and ID numbers are redacted from payloads.
Agreements: DPA and NDA on request. Regional rules covered in our privacy policy.

Keys and sensitive data

How API keys, tokens and personal data are handled.

Keys are kept in a vault
API keys and tokens are stored in the orchestration layer's encrypted credential store or the platform's secret manager. None are written into workflows, repos, spreadsheets or chat.
One credential per connection
Each integration gets its own key, scoped to the objects and actions it needs. A lead sync cannot read invoices. Revoking one key stops one thing.
Service accounts for connections
Connections use a dedicated service user in your account, so nothing breaks when a person leaves and nothing runs under someone's name.
Two-factor on every account
Each platform account we set up has two-factor authentication on, and the recovery codes are handed to you.
Rotation on handover and on request
Keys are rotated when the build is handed over, when anyone with access leaves, and at any time you ask.
Secrets are redacted from logs
Audit logs redact tokens, card numbers, government IDs and anything you flag as sensitive before a line is written.
Webhooks verified before processing
Inbound calls are checked against the sending tool's signature. Unsigned or stale payloads are dropped and counted.
Access you can see and revoke
Our access is through managed roles in your accounts. You can see it, limit it or remove it without touching the build.

Engineering decisions

Decisions we make before writing the first workflow.

Rate limits and backoff
Each API has a ceiling, so calls are batched, queued and retried with increasing delays during busy periods.
Backfill and pagination
The first run loads history as well as new events. Backfills page through the source in order and resume after an interruption without writing a record twice.
Out-of-order and duplicate events
Webhooks can arrive twice or out of sequence. Each event carries a version or timestamp, so an older version cannot overwrite a newer one.
Schema changes
A renamed field or a new dropdown value is a common cause of silent failure. Payloads are checked against a contract, and a mismatch stops the run and raises a flag.
Time zones, currencies, units
Dates are stored in UTC and shown in local time. Amounts carry their currency and quantities their unit, and conversions happen once, in the map step.
Sandbox first
Builds run against sandbox or test accounts with seeded data first. Going live means switching credentials.
Versioned and reversible
Workflows and code are version-controlled. A release can be rolled back in minutes without losing the records processed since.
API deprecations watched
Tools retire API versions on a schedule. We track the dates for every connection in the runbook and upgrade before the cut-off.
Cost per run known
Each run logs its cost, including any model calls, so you can see spend as it builds up.
Run metrics
We chart runs per hour, records in and out, exception rate and latency, which often shows a problem before an alert fires.

Proof

Where we have done this.

Customer story
Enquiry and quotation platform on SAP

An enquiry platform for an industrial distributor, reading emailed RFQs with AI and matching parts to SAP, with OpenAI and Retool in the stack.
Read the story

The offer

Hire SBD to connect OpenAI and Retool.

We agree a fixed scope in a discovery call and go live in three to five business days. The accounts are in your name, and the engineers who built it stay on call for sixty days after.

Hire SBD for this integration See similar integrations

ScopeOne call to agree the flows, the fields and the owner, written up the same day.
Build4 workflows, the data map above, validation, upserts, retries and reconciliation.
Go-liveWe test in a sandbox first, then switch over. Day three to five from scope for most pairs; longer only if a tool needs approval, such as WhatsApp templates.
HandoverRunbook, training session, alerts to a named owner, admin access and code in your accounts.
AfterSixty days of hypercare, then a care plan or run it yourself.

Questions

What people ask about this pair.

Will OpenAI train on our data?

OpenAI states that business API data is not used for training by default. We confirm the current terms for your account and limit what is sent.

Could we use Claude instead?

Yes. We keep the model step easy to replace, and test candidates on a sample of your real documents.

How accurate will it be?

We measure it on your documents before launch, and keep a person on anything uncertain.

Keep exploring

Related to this page

OpenAI and Retool