Skip to content
Splice Work

AI integration

AI integration solutions for the software you already use

Most AI proposals start by asking you to move. This one does not. The systems stay where they are, and the agent goes to them.

Book a walkthrough

The argument

You do not need new software.

The expensive part of most AI projects is never the model. It is the migration underneath: a new system of record, a data cleanup, months of change management, and a team that quietly keeps using the old spreadsheet anyway. By the time anyone measures the result, the thing being measured is the migration.

An integration inverts that. Your CRM stays your CRM. Your ledger stays your ledger. The agent is given scoped credentials to the systems you already pay for, and it does its work inside them. If it turns out not to earn its keep, you switch it off and nothing about how your business runs has changed. That is a much cheaper thing to be wrong about.

Three patterns

Almost every integration is one of these three.

Pattern 01

Read, decide, write back

The workhorse. The agent pulls a record on a schedule, works out what should happen to it, and puts the answer back in the same system. Nothing new to open, nothing to check.

  1. Step 1Read the recordOpen invoices and their aging, straight out of the ledger.
  2. Step 2DecideWhich are genuinely overdue, which are disputed, which were already paid.
  3. Step 3Write it backDraft the follow-up, log the activity, update the status.

Invoice follow-up running on QuickBooks Online and Salesforce.

Pattern 02

Event, agent, notify

For work that cannot wait for a schedule. Something happens in one system, the agent handles it immediately, and a person hears about it only when they need to.

  1. Step 1Something happensA call comes in after hours, or a payment posts in Stripe.
  2. Step 2The agent actsAnswers, or stops the collections sequence already in flight.
  3. Step 3Tell the right personA card in Slack with what happened and what it did.

After-hours answering on Twilio, with the summary posted to Slack.

Pattern 03

Document in, data out

The one that replaces re-keying. A document arrives in a format nobody can query, and the agent turns it into structured records in the system that should have had them all along.

  1. Step 1A document arrivesA PDF purchase order, a supplier invoice, a signed quote.
  2. Step 2Extract and checkPull the line items, match them against the order they belong to.
  3. Step 3Post itCreate the record, or route the mismatch to a person instead.

Bills landing in BILL, matched to their purchase orders before anyone approves them.

The stack

The systems we already have working.

  • Salesforce
  • QuickBooks Online
  • BILL
  • Stripe
  • Airtable
  • Slack
  • Microsoft Teams
  • Google Workspace
  • Twilio

These are the ones running in production today, not a compatibility list. If yours is not here and it has an API, connecting it is a known quantity rather than a research project.

Go deeper

What an agent reads in each system, and what it is allowed to write back, is set out system by system.

See every integration

Choosing someone

What to ask an AI integration company before you sign

Classic system integration moved records between two applications on a schedule and called it finished. AI system integration adds the step in the middle where something has to be judged, and that step is where the projects go wrong. A traditional AI systems integrator will scope the pipes accurately and leave the judgment undefined.

Most of what is sold as AI integration for business is a pilot with no route into production: a demo on sample data, a report, and an invoice. The questions on the right are the ones that separate that from work that ships. Ask them of us too.

Our own answer is narrow on purpose. We build one agent for one job and run it, rather than selling a programme. Where the job needs more than an integration — a system built from scratch — that is a different conversation, and we will say which one you are in.

  1. 01Which of the three patterns is this, and why that one?
  2. 02Whose credentials does it run on, and scoped to what?
  3. 03What happens the first time the model is wrong?
  4. 04What do we keep if we stop working with you?
  5. 05Who picks up the phone when it breaks on a Sunday?

The part people ask about last

What happens to your data

It stays in your systems

There is no Splice Work database that becomes the new home for your records. The agent reads what it needs at the moment it needs it and writes the result back where it belongs.

It is not used for training

We run on the business API tiers from the model providers, which do not train on content sent through them. Your customer list does not end up improving somebody's model.

Access is yours to revoke

Every credential is issued by you, scoped to the narrowest thing that lets the job finish, and documented. Turning an agent off is a permissions change you can make without us.

The logs are the other half of it. Every action an agent takes is recorded with what it read and why it acted, which is the difference between a wrong answer you can fix and one you can only apologise for. Those logs are yours to read.

Questions we get asked

What does AI integration actually mean here?

Connecting a language model to the systems that already run your business, with permission to read specific data and write specific things back. It is not a chatbot bolted onto your website. The measure of it is whether work finishes without a person re-typing anything.

Do we have to replace any of our software?

No, and if a proposal starts with a migration you are buying a platform, not an integration. The whole approach assumes Salesforce stays Salesforce and QuickBooks stays QuickBooks. The agent works inside them.

How long does an integration take to stand up?

It depends almost entirely on how clean the access is. A documented API with credentials ready is a different job from an on-premise system with no API. We scope it after looking at the actual systems, not before.

Is our data used to train an AI model?

No. We run on the business API tiers from the model providers, which do not train on the content sent through them. Your records also stay in your systems rather than being copied into a database of ours.

What can the agent see?

Only what the credentials you issue allow. We ask for the narrowest scope that lets the job finish, we document what each one is for, and you can revoke any of it without our involvement.

What happens if the model gets something wrong?

Financial and customer-facing actions wait for a human approval, so the mistake is caught before it lands. Every action is logged with what the agent read and why it acted, which is what makes a wrong answer fixable rather than mysterious.

Can we start with one workflow?

That is the only way we recommend starting. One job, running in production, earning the right to a second. Buying an AI programme before a single agent has worked is how these projects die.

Bring the workflow that crosses two systems.

Thirty minutes on your actual stack. You will leave knowing which of the three patterns it is, and what it would take.

Book a walkthroughSee the integrations