Skip to content

Integration: Artificial Intelligence (AI)

AI reads your documents, sorts your requests and answers from your own data, inside the software your teams already use. The hours spent typing and searching go back to the work that matters. Every result shows its source or how confident it is.

Talk to an expert

For
SMBs and organizations handling lots of documents or requests
Technologies
Laravel, Vue, OpenAI, Claude, Gemini, open-source models
Team
Montreal, all across Quebec

01 When

When AI belongs in your operations

AI excels at tasks that come back every day, follow known rules and can be checked at a glance. If you recognize one of your tasks below, there are hours to win back.
A team working together in front of several screens of code
  1. Documents retyped by hand

    Invoices, purchase orders, forms: they arrive as PDFs, photos or spreadsheets. Then someone types them into your system.

  2. Requests to sort before anyone can act

    Every morning, someone reads the incoming emails and forms just to work out who should handle them.

  3. The same questions, every week

    The answer is already in a policy, a contract or a file. You just have to know where to look.

  4. Judgment that follows known rules

    Your team can explain why a request goes here rather than there. Those rules fit in a few sentences.

  5. Volume growing faster than the team

    Documents and requests grow year after year. AI absorbs the increase, and your team keeps its time for work that takes judgment.

We first find where AI will have the biggest impact on your operations: the places where it gives your team back the most hours. That is where we start.

We force nothing; return comes first. Where a simple rule or an existing product does better, we tell you.

02 Uses

What AI does inside your software

It works in your screens, on your data, and takes on the reading, the sorting and the first draft. Your team keeps the decisions. How our developers use AI to build faster is covered on our custom software page.

02

Sorting and routing

Every request, email or photo is sorted and sent to the right person as soon as it arrives, with a confidence score. Anything uncertain goes to review.

See OrionFlow : Sorting and routing

04

Drafting

A first draft of a reply, a summary or a letter, written from your data. Your team edits it instead of starting from a blank page.

See Matchwork : Drafting

05

Search by meaning

The right passage found in your documents, even when the words you typed are not in it. Everyone sees only what they already have access to.

Read the article : Search by meaning

03 Method

From the first meeting to launch, every step leaves a record

The method behind all our projects, applied to AI: we start from your real examples and put in writing what the AI is allowed to do.

01 Discovery

We start from your real documents.

Meetings with the people who do the task today. Together we gather real examples, including the hard ones: a badly scanned page, an ambiguous request.
Fig. 01 The discovery meetings board

02 Analysis

We trace where every piece of data goes.

Where the document arrives, where it must end up, who is allowed to see it. With you, we decide what can be sent to an AI model and what never leaves your systems.
Fig. 02 The map of tools and data

03 Data model

Every AI proposal has its place in the database.

The model answers in a strict format, field by field. We keep the proposed value, its confidence score, the value kept and the person who approved it.
Fig. 03 The data model where every proposal waits for approval

04 Mockups

The review screen is designed first.

The queue of documents to review and the panel for each one, with the fields the AI read. Your team comments on it before a line of code is written.
Fig. 04 The wireframe of the review screen

05 Specification

We also write down what the AI is not allowed to do.

The specification sets the thresholds, what goes back to a person and what gets logged. You approve it before development starts.
Fig. 05 A page of the specification

06 Development

A version to try on your examples, every week.

Whatever is finished during the week lands in staging. You run your own documents through it, then we go over it at the follow-up meeting.
Fig. 06 This week’s board

07 Quality

The safeguards are tested like the rest of the code.

Code review by a senior developer, automated tests and staging. The tests also cover the rules around the AI: a field below the threshold goes to review, a correction is never overwritten.
Fig. 07 The review threshold, from code review to production

04 Control

Your rules decide what a person checks

The AI does the first reading. Your team sees where each answer comes from and what it is unsure about, then they decide. The two screens below are examples of what we build, with made-up data.

Assistant

An answer that cites its sources.

The assistant answers in seconds from the documents the user is allowed to open and points to the exact passage. If the answer is not there, it says so.

Extraction

Below the threshold, a person checks.

The invoice on the left, the fields read on the right. The purchase order and the category fall below the 80% threshold: the invoice waits for a person to check them before it goes to accounting.


Choices that make AI dependable day to day

Choices taken from software in production, such as OrionFlow and the Observatoire des tout-petits data platform. Read the OrionFlow case study

  1. A confidence threshold you set

    The risk
    Rereading every result wipes out the time saved. Rereading none lets mistakes reach your customers.
    Our choice
    Every result carries a confidence score. Below the threshold you set, it waits for a person in a separate queue.
    The gain
    Your team only checks what needs checking.
  2. A human correction stands

    The risk
    A fresh AI reading overwrites a person’s work.
    Our choice
    Whatever a person has sorted or corrected is never sent back to the AI. When the stakes call for it, nothing is written to the database until someone approves it.
    The gain
    Approved work stays approved.
  3. One outage does not block everything

    The risk
    An outside AI service has outages and limits. If everything moves as one batch, a single failure stops the day.
    Our choice
    Each document follows its own path. A failure is retried later, a few times at most, then shown with its reason.
    The gain
    The rest of the work carries on. You can see what is waiting and why.

05 Data

Your data and your costs, under your control

What may leave your systems is decided with you, before the first line of code. For the rest, here is what we can put in place to suit your needs.

Your data

  • We choose with you which fields are sent to the model. The rest never leaves your systems.
  • When needed, personal information is stripped from what goes to the model, if the task can do without it.
  • We use business access whose terms rule out training on your data. You choose the provider with us, terms in hand.

Quebec’s Law 25

  • People must be told when a decision about them rests exclusively on automated processing of their personal information.
  • They can learn the main factors behind the decision and submit their observations to someone who can review it.
  • The software can keep a person in the loop and log what the AI proposed and what was decided.

Your costs

  • The cost of model calls can be logged per feature, so you know what each use costs you.
  • A daily or per-feature cap can limit the provider’s bill.
  • You can hand simple tasks to a cheaper model and keep the most capable one for work that needs it.

Your results

  • Before building, we pick a few measures with you: time per document, errors per field, share sent for review.
  • We take them on your real examples, before and after rollout.
  • The gain shows up as figures on your own documents, not as a general promise.

Under the hood

The building blocks of an integration, chosen to last and to switch models without starting over. See our technologies

Models
OpenAI, Claude (Anthropic), Gemini, open-source models, chosen for the task, the cost and how sensitive the data is
Application
Laravel, Vue, AI built into your screens
Processing
Queues, each document on its own, retried on failure
Answers
Strict, validated format, every field checked before it is saved
Integrations
APIs, webhooks, email, ERP, CRM, accounting software
Log
Request, answer, confidence, decision, kept as your traceability needs require

06 Rollout

From first trial to production, one step at a time

We start small, on your documents, and widen the scope as soon as the results justify it. The prototype shows you the real gain before you invest in the full integration.
  1. Real examples Your documents and requests, hard cases included.
  2. Prototype Within a few weeks, the AI tried on those examples.
  3. Decision Go on, adjust or stop, with the results in hand.
  4. Integration The AI wired into your systems and screens.
  5. Pilot group A few users, every result checked.
  6. Production Thresholds are adjusted as trust builds.

07 After

After launch, we build on what pays off

AI models improve every few months. We follow what the AI achieves in your software, adopt new models when they do better and find with you the next task to hand it. See also our support and maintenance offer.

01

Correction reviews

At an agreed interval, we go over the cases your team corrected. They show where to adjust an instruction or a threshold.

02

New models

A more accurate, faster or cheaper model comes out regularly. We try it on your examples and adopt it when it does better.

03

Settings in your hands

Thresholds, instructions and categories become settings your administrators change without a developer.

04

The next task

Another document type, another queue of requests: we pick with you the one where AI would pay off most. The same mechanism applies: intake, reading, review, log.

08 Work

AI already at work for our clients

Each case study shows the work the AI takes on and the part the team keeps.

View all projects

  • OrionFlow

    Contents claims

    A whole house in photos. AI makes the inventory.

    Custom software

  • Québec’s youngest children. In over 200 indicators.

    Custom software

  • Workstart

    Tech startup

    Candidates answer. Agencies know who to call.

    Custom software

  • Boursify

    Financial education

    A market that doesn’t exist. Whole classes trading live.

    Web application, Mobile application

Our experience with Witify has been nothing short of exceptional. [...] Their consistently innovative and effective results have earned our highest recommendation for any company looking for a superior long-term technology partner.

Kenya Kondo President, Workstart

09 Questions

Questions and answers

What we are asked before a first AI project.

Another question? Email us at info@witify.io or call 1 800 334 9031.

How can AI really help a non-tech SME?

AI adapts to your specific needs. You do not need technical expertise: we start from your documents, your requests and the way your team handles them today. The AI then takes on the repetitive tasks, such as reading an invoice or sorting an email. Your team keeps the decisions.

We support you every step of the way. We start with the task where AI will save you the most time, and if a simple rule is enough elsewhere, we will tell you.

How long does it take to deploy an AI solution?

The duration depends on the complexity of the project, the amount of data available, and your existing environment. An initial prototyping phase can be completed in a few weeks, while a full deployment can span a few months.

After the prototype, you decide to go on, adjust or stop, based on measures taken on your own examples.

Is artificial intelligence safe for my data?

We pay particular attention to data confidentiality and security. We decide with you which fields can be sent to an AI model. When needed, personal information is stripped out. Model calls can also be logged.

We use business access whose terms rule out training on your data. You choose the provider with us, terms in hand. The software and its database are hosted on ISO 27001 and SOC 2 certified servers.

What return on investment can I expect?

Since every project is unique, ROI will vary depending on your goals and the scope of the solution. The gain mostly comes from the reading and typing time taken off your team, and from errors caught before they get expensive.

To put a number on it, we agree on a few measures before building: time per document, errors per field, share sent for review. We then take them on your examples, before and after. Either way, the prototype tells you early whether the gain is there.

Can we integrate AI with our existing systems (ERP, CRM, etc.)?

Yes. AI does not replace your systems: it plugs into them. It reads what comes in (emails, shared folders, exports), prepares its proposals in your screens, then writes to your ERP, CRM or accounting software once the result is approved.

Our team knows systems integration well. Wiring into your tools often takes more work than calling the AI model itself.

What if my needs change after deployment?

We offer continuous support and an agile approach. Thresholds, instructions and categories are settings you change yourself. For the rest, we adjust the solution with you: a new document type, a new task or a newer model, which we can try on your examples first.

What happens when the AI gets it wrong?

Like a person, AI can make mistakes. Its strength is that it states how confident it is, and the software builds on that: every result carries a confidence score. Below the threshold you set, a person checks before the data goes any further. A well-designed assistant cites its sources and says when your documents do not hold the answer.

We also design the software so that a correction made by hand is never overwritten by a fresh reading. The success rate is measured on your own documents, from the prototype on.

Why not just give ChatGPT or Copilot to our team?

To draft an email or summarize a text, a consumer tool does the job well: we will tell you so. Integration makes sense when the task repeats, touches your data and has to leave a record: reading every invoice received, sorting every request, answering from your internal policies.

Inside your software, the AI only sees what the user is allowed to see. It writes to your systems once the result is approved. The software keeps a record of what it proposed.

Does Quebec’s Law 25 apply to an AI project?

As soon as a project processes personal information, yes. The law also sets rules for a decision based exclusively on automated processing: the person concerned must be informed. They can learn the main factors behind the decision and submit their observations to someone who can review it.

We can design the software to keep a person in the loop where it matters, with a log of what the AI proposed and what was decided. For your specific situation, your legal counsel remains the reference.

How much does it cost to integrate AI into software?

There are two costs. The first is development: a prototype on your examples is a bounded first step, estimated after the discovery meetings. Our article on what custom software costs explains what moves a budget.

The second is usage: every call to the model is paid to the provider. We can track it per document and cap it, so it stays predictable.

A task that looks made for AI? Let’s look at it together.

In 30 minutes, an expert goes over a few of your documents or requests. You will know where AI would have the biggest impact and where to start.

  • A 30-minute call with an expert, not a salesperson
  • A realistic budget estimate, backed by numbers
  • A concrete action plan, whether you work with us or not
Or write to us

Free. No commitment. Reply within 24 hours.

Names and figures shown in the interfaces are examples.