Most artificial intelligence projects don’t fail for lack of technology: they fail because nobody uses them. A tool is launched, the demo is given, there’s enthusiasm for two weeks and then the team goes back to its usual way of working. That’s why the case of Fyxer is interesting, a company that has built an AI assistant for management tasks —email, meetings, follow-up on commitments— and that publishes two figures that speak of real adoption, not expectations: 53% of the drafts it generates are sent as they are and more than 90% of its users are still paying for it and using it after 90 days (source: OpenAI case study, How Fyxer built an AI executive assistant people trust, openai.com/index/fyxer).
We’re not going to retell their story. What’s interesting is something else: which design decisions explain that trust and how they translate to a Spanish private company considering its first AI project.
The real problem: what’s easy for a person is hard for a machine
Two people can receive the same email and need completely different replies, depending on the relationship with the sender, what was discussed three weeks ago and each one’s objective. A good human assistant solves this without thinking; a language model, without context, writes something correct and perfectly useless.
It’s a situation that anyone who manages a company will recognise. Critical knowledge —which client will tolerate a delay, which supplier has to be chased, which order isn’t shipped without credit confirmation— isn’t written down anywhere. It lives in the heads of three or four people. Any AI project that ignores that context produces generic results, and generic results don’t get used.
Four keys to building an AI assistant that people actually use
1. Break the task down into small decisions
Fyxer didn’t ask a single model to "write a good email". It broke the work down into dozens of specialised models, each responsible for a narrow part: does this require a reply or just information?, where is the conversation heading?, which details from previous exchanges are relevant now?
The lesson is transferable and very practical: instead of framing it as "we want AI for the sales department", it’s better to isolate concrete, measurable decisions. Classifying order emails. Extracting data from a delivery note. Detecting which quotes deserve follow-up this week. Each piece can be measured separately and, if it fails, you know exactly where.
2. Train with trade knowledge, not with theory
Before selling software, Fyxer ran a human assistant service. That’s where its advantage came from: a dataset built on more than 500,000 hours of real annotated work, capturing the small judgement calls of a good assistant —when to reply quickly, when to wait, which previous conversation matters.
Translation for an SME: your competitive advantage in AI isn’t in the model, which is the same one anyone can access. It’s in your data and in your operational judgement. The ERP’s history, the unwritten rules of your warehouse managers or your site manager, the way your sales team qualifies an opportunity. That’s where a project becomes hard to copy.
And that’s also where the uncomfortable part lies: if that data is scattered across spreadsheets, duplicated or ownerless, AI doesn’t perform. That’s why at Tisa we start with the boring part —identifying which data is relevant and putting data governance protocols in place— before talking about models.
3. Turn user corrections into learning
When someone edits a draft before sending it, they’re giving a free lesson: the difference between what the machine proposed and what the person sent is the most honest signal there is. Fyxer captures that difference and uses it to improve the system, and every change goes through an A/B test before deployment: it’s only released if it improves results in a statistically significant way.
For a company of 50 or 200 employees, the realistic version of this isn’t a continuous retraining pipeline. It’s something more modest and perfectly achievable: a channel where the user flags "this is wrong", someone who reviews those corrections every week and a routine for adjusting the system. Without that loop, the tool freezes on launch day and loses credibility at the first mistake.
4. Measure adoption and trust, not just enthusiasm
Fyxer’s co-founder puts it well: everyone talks about revenue, but retention is the real test. It’s worth copying that criterion when evaluating any AI pilot.
Before starting, define two or three indicators: percentage of proposals accepted without changes, minutes saved per file, errors detected before invoicing, active users after 60 days. If after three months nobody uses the tool, the project has failed, even if the technology works flawlessly. A measurable use case is the difference between a project and an anecdote.
How we bring this down to earth at Tisa
Our artificial intelligence and data line works in levels, according to each company’s maturity, and always with the business ahead of the technology:
- AI advisory: identifying the relevant datasets, evaluating tools according to real investment capacity, defining data governance aligned with regulations and training the team.
- Data analysis: from putting information in order and visualising it with Power BI to working with multiple sources to decide with data rather than by intuition.
- Development and execution of a use case tailored to the business: a specific problem, a defined scope and metrics agreed before starting.
All of that fits on the foundation we already know: Microsoft Dynamics 365 Business Central as the management system and Power Platform to automate and extend processes. We’ve been implementing solutions since 1987 in retail, distribution, construction, the food industry, hospitality, sports centres and fleets, and it’s that sector experience that makes it possible to tell which automation adds margin and which only adds noise. You can see some examples in our success stories.
An honest warning: not everything announced today is mature. Assistants built into the ERP itself are advancing fast, but it’s advisable to verify specific capabilities in each version before building a critical process on top of them
Where to start
If you’re considering your first AI project, the order we recommend is simple: choose a process that genuinely hurts, check the state of the data that feeds it, define a pilot with metrics and decide after three months whether it scales or is dropped. A useful AI assistant isn’t born from a big rollout, but from a small problem solved very well.
Would you like a no-obligation assessment of your case? Let’s talk: (+34) 971 305 885 · info@grupotisa.com · grupotisa.com.
Source of the case discussed: OpenAI, "How Fyxer built an AI executive assistant people trust" (openai.com/index/fyxer). The figures cited are from that publication and not from Tisa projects.
