On 17 September 2026, OpenAI unveiled Astra for Law, a configuration of its most powerful model designed specifically for legal work. At first glance it is news for law firms. But if you run a distribution company, a hotel group or a construction firm, the announcement matters to you for another reason: it is the clearest demonstration to date of how vertical AI is built, that is, artificial intelligence fine-tuned for a specific trade. And that recipe can indeed be applied, on another scale, in your business.
What OpenAI announced (and what lies beneath the headline)
According to the company itself, Astra for Law is not a new model built from scratch: it is its latest-generation model combined with three things: a legal search index, specific instructions for legal analysis and drafting, and privacy and governance controls for confidential work. OpenAI states that the index covers United States case law, regulations and rulings over a corpus of more than 230 million URLs, and that in a legal research benchmark the combination notably outperforms the same model using only web search.
Source: OpenAI, «Introducing Astra for Law» (openai.com/index/astra-for-law, 17 September 2026).
What matters is not the legal sector. It is the pattern. The generic model was already good; what turns it into a reliable professional tool is the context surrounding it.
The four ingredients of a vertical AI
When at Tisa we analyse AI projects applied to business, the scheme that works looks very much like this one. A useful vertical AI needs four pieces:
1. A good model, but as a starting point
The model is the engine, not the car. Choosing between providers (Microsoft/Copilot, OpenAI, Anthropic, Google or open options) matters, but it is the least differentiating decision: models change every few months and they all advance. What you build around them is what lasts.
2. Your data, accessible and organised
The legal index of Astra for Law is, in essence, a reliable knowledge base that the model consults before answering: the technique that in the industry we call RAG (retrieval-augmented generation). In an SME that "index" is not case law: it is your ERP data, your price lists, your supplier contracts, your technical datasheets, your service histories. Without that anchoring, a model answers with plausibility but without truth; with it, it answers with sources that a person can check.
3. Instructions that encode professional judgement
OpenAI adds instructions on how to reason and write a legal document: distinguishing the essential from the incidental, flagging weaknesses, explaining risks. Translated to your company: how a special order is assessed, which margins can be moved, what an experienced site manager warns about before accepting a change. That knowledge is usually in the heads of three or four people. Turning it into explicit instructions is, very often, the most valuable part of the project.
4. Confidentiality, permissions and oversight
The announcement devotes a whole section to control: zero data retention, exclusion from human review, information permissions and internal oversight. In a private company the equivalent is just as serious: who can ask what, what data leaves the organisation, what is logged and who validates before an answer becomes a decision. This is data governance, and it has been part of our AI advisory work from day one.
From the law firm to your company: how it translates
The useful question is not "when will there be an Astra for Law for my sector?", but "which specific task in my company improves with a vertical AI built on my data?". Some recognisable examples in the sectors we work with —retail, distribution, construction, food industry, hospitality, sports centres and fleets—:
- Distribution: an assistant that answers "what did we offer this customer last year and at what margin?" by querying the ERP, instead of forcing people to dig through reports.
- Hospitality: occupancy and rate analysis that cross-references bookings, allotments and agency invoicing to anticipate the coming week.
- Construction: assisted review of tender documents and subcontractor agreements that flags deadlines, penalties and deviations from what is usual in the company.
- Industry and fleets: detection of patterns in incidents and maintenance to get ahead of breakdowns and properly allocate cost per vehicle or line.
- Retail: purchasing support with turnover history and seasonality, integrated with inventory.
None of these cases needs a proprietary model. It needs organised data, integration with the management system and business judgement.
The ERP as the foundation of vertical AI
Here is the least glamorous and most important conclusion: without reliable data there is no AI worth having. An ERP such as Microsoft Dynamics 365 Business Central concentrates finance, purchasing, inventory, sales, projects and service in a single place, with consistent and traceable information. That is exactly the material a vertical AI needs in order to stop improvising. If the information lives scattered across spreadsheets, emails and an old system, any AI project will start there anyway.
That is why, in many of our projects, the first step towards AI is actually a step towards order: consolidating the ERP, connecting what was loose and setting up dashboards with Power BI. Afterwards, the AI layer delivers much more and costs much less.
How to start without burning through your budget
Our recommendation is the same as always, and OpenAI’s announcement reinforces it: start with a measurable use case, not with the technology. A practical script:
- Choose a task that is repetitive, involves expert judgement and has identifiable cost (hours, errors, delays).
- Review the data that task needs: where it is, who maintains it, whether it is reliable.
- Define the rules of the trade and the limits: what the machine decides and what a person always validates.
- Pilot small, measure against the current way of working and decide with data whether to scale up.
- Train the team, because a vertical AI without trained users is abandoned within three months.
With that approach, the conversation stops being about fashions and becomes about margin, time and errors avoided.
Let’s talk about your case
At Tisa we have been implementing management technology since 1987 and, as a Microsoft partner and ISV, we have developed industry verticals on Business Central for hotels, distribution, sports centres, point of sale and fleets. That same logic —knowing the trade, not just the tool— is what we apply today to AI projects.
If you want to assess which vertical AI use case makes sense in your company and what you need before tackling it, write to us at info@grupotisa.com or call us on (+34) 971 305 885. First assessment with no obligation, and we will tell you frankly whether it is time for AI or time to put the house in order first. More information at grupotisa.com.