Table of Contents
Write an industrial AI RFP around the problem, not the technology. Describe the decision you want to improve and today's baseline, document your data honestly and share a representative sample, define the acceptance test and pass mark, publish the scoring criteria and weights, and require ownership, knowledge transfer and exit terms. Ask for a price per phase.
Key takeaways
- Specify the problem and the performance to reach, not a product. Algeria's Law 23-12 requires public buyers' technical specifications not to be oriented towards a given product or supplier.
- Give every bidder the same representative data sample at the same time, under written conditions on use and deletion.
- Put the acceptance test in the RFP: metric, baseline, test data the bidder has never seen, and pass mark.
- Publish the scoring criteria and their weights. For buyers under Law 23-12, this is mandatory.
- Ask for a price per phase. Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025 (Intelligent CIO).
Why does an AI project need a different RFP?
A standard IT tender assumes you can describe the system and check it on delivery. AI differs in three ways. Performance depends on your data, so nobody, including the vendor, knows the result before testing on it. You supply part of the input: historian tags, lab results, maintenance records, manuals. And accuracy changes as the process changes, so the contract must cover years of operation, not one handover.
The World Economic Forum's AI procurement guidelines, developed with fellows from the UK government's Office for AI, note that AI systems "learn and adapt through ongoing, periodic training with new data" and recommend a process that allows room for iteration.
This guide covers the document you send. For the questions to put to vendors and a sheet to score their answers, see our vendor selection checklist.
What sections should an industrial AI RFP contain?
| Section | What to include |
|---|---|
| 1. Context and problem | The decision to improve, who makes it, today's baseline |
| 2. Scope and constraints | Sites, users, languages, where the system must run (cloud, on premise, offline) |
| 3. Data annex | Systems, volumes, periods, formats, known gaps, sample data |
| 4. Functional and performance requirements | What the system must do and the level to reach |
| 5. Acceptance tests | Metric, baseline, test data, pass mark, duration |
| 6. Security and integration | Network zones, interfaces, SBOM, AI-specific testing |
| 7. Deliverables and knowledge transfer | Documentation, training, handover |
| 8. Ownership, licences and exit | Data, fine-tuned models, outputs, export at the end |
| 9. Price schedule | Price per phase and per year of operation |
| 10. Evaluation and response format | Criteria, weights, evidence required, deadlines |
How do you write the problem statement?
The WEF guidelines advise to "focus on developing a clear problem statement, rather than on detailing specifications of a solution". The UK guidelines say: "Tell suppliers about the situation or challenge, and let them propose a solution that meets your needs."
Compare two versions:
- Weak: "Supply an AI platform based on a 70-billion-parameter language model and a vector database."
- Better: "Maintenance technicians at our two sites search French and English pump manuals to diagnose faults. We want answers drawn from these documents, citing the page, with documents kept on site. We will measure the share of correct, cited answers on real technician questions."
State the baseline: how long a diagnosis takes now, or how much off-spec product is made per month. Without it, nobody can measure an improvement.
Leave room for answers without AI. The WEF toolkit says that if another solution "that doesn't incorporate AI is better able to address the problem", you should follow that path. Our article on why industrial AI projects fail explains why a vague problem sinks so many projects.
For public buyers in Algeria, this is also the law. Article 16 of Law 23-12 requires needs to be set out in detailed technical specifications based on standards and/or performance or functional requirements, which "must not be oriented towards a given product or economic operator" (our translation).
What should the data annex contain?
Unless you tell them what exists, bidders price data work by guessing. For each source, list the system (historian, LIMS, maintenance management system, document repository), period, volume, format, languages and share of scanned documents. Say what is missing: tags renamed after a revamp, lab results on paper, failures never recorded. The WEF toolkit recommends highlighting known data limitations and requiring "tenderers to describe their strategies on how to address these shortcomings".
Then share a sample. The WEF guidelines suggest releasing data during procurement so bids match your situation and, for sensitive data, only a sample representative of the whole dataset. The UK guidelines add: "Ensure you provide the same data to all suppliers at the same time." Following the WEF toolkit, write down:
- minimum requirements for the environment where the bidder holds the data;
- a signed commitment to use the data only for the bid;
- a deletion date and written confirmation of deletion.
Check for personal data first, since maintenance reports and shift logs often name employees. In Algeria, Law 18-07, amended by Law 25-11 of 24 July 2025, subjects transfers of personal data abroad to prior authorisation from the national authority, the ANPDP, save for exceptions in the law (Gide). Remove names and confirm with counsel before sending data to a foreign bidder. If your documents mix Arabic, French and English, include all three; see handling multilingual documents.
How do you define acceptance tests?
Write the test before you receive the bids:
- A metric linked to the decision. For a document assistant, correct cited answers; for a soft sensor, error against the lab; for predictive maintenance, warnings before confirmed failures and false alarms.
- A baseline. Today's performance on the same metric, measured the same way.
- Test data the bidder never sees. Hold part of your data back and run the test yourself.
- A pass mark and a duration. For example, a pilot covering normal operation and at least one grade change or maintenance stop.
- A test environment. The joint guidance led by CISA recommends testing AI systems first on infrastructure designed for testing, then on more realistic non-production systems (CISA et al.).
- Monitoring after acceptance. The UK guidelines state that "testing the model on an ongoing basis is necessary to maintain its accuracy." Specify who measures performance after go-live, how often, and what triggers retraining.
For LLM projects, see evaluating an LLM on your own documents. Where the AI recommends actions, state where a person approves them; see human-in-the-loop AI agents.
What security and integration requirements belong in the RFP?
The CISA-led guidance, published in December 2025 with partner agencies including Germany's BSI and the UK's NCSC, lists what operators should demand from vendors:
- explicit contractual details on AI features;
- a software bill of materials (SBOM), and where AI models are hosted;
- notification if the vendor finds the AI can give improper advice or take inappropriate action;
- a data usage policy covering residency, communication paths, encryption and storage, including whether the vendor may train on your data;
- the ability to operate on premises or without continuous internet or vendor cloud access;
- operator control over enabling and disabling AI features.
For LLM systems, ask for a test report against the OWASP Top 10 for LLM Applications 2025, especially prompt injection (LLM01), sensitive information disclosure (LLM02) and excessive agency (LLM06). Ask which framework bidders follow, such as the voluntary NIST AI Risk Management Framework or ISO/IEC 42001, and for the scope of any certificate. For network zones and interfaces, see integrating AI with SCADA and OT systems and securing an on-premise LLM.
What should the RFP say about ownership and exit?
- Documentation. Require the equivalent of a model card: a short document stating intended use and how the model was evaluated.
- Ownership. The UK guidelines advise considering "whether you or the supplier should own any new intellectual property". Settle it for your data, fine-tuned models, configurations and outputs.
- Lock-in. The WEF toolkit recommends requiring interoperability and "open licensing terms to avoid vendor lock-in".
- Knowledge transfer. The WEF toolkit says to "make knowledge transfer a requirement under the RFP". For Algerian public tenders, Article 60 of Law 23-12 requires the specifications to provide, as applicable, for training and the transfer of technology and know-how.
- Exit. Specify export formats for data and models, the cost of exit, and deletion of your data by the vendor.
For clause wording, the EU's community of practice on public procurement of AI publishes model contractual clauses for AI, updated on 5 March 2025, in a full version for high-risk AI and a light version for other AI. They follow the EU AI Act, so adapt them with counsel.
How should you score and price the bids?
Publish the criteria, weights and evidence expected. Under Article 53 of Law 23-12, criteria and weights must be linked to the contract's object, non-discriminatory and stated in the specifications, and the technical evaluation must suit the project's nature, complexity and importance. Article 54 forbids negotiation with bidders in a tender, so the RFP must be complete when published.
An example grid to adapt:
| Criterion | Example weight | Evidence to request |
|---|---|---|
| Understanding of the problem and approach | 20 | Method note tied to your baseline and data |
| Results on the sample data | 20 | Results on the shared sample, with the method |
| Data, hosting and security | 15 | Architecture diagram, data usage policy, SBOM |
| Integration with your systems | 10 | Named interfaces, effort in person-days |
| Team and references | 10 | CVs of named staff, production references |
| Ownership, knowledge transfer and exit | 10 | Proposed clauses and training plan |
| Price over three years | 15 | Completed price schedule |
| Total | 100 |
Score with a mixed panel. The WEF toolkit notes that "technical, business as well as legal and ethical experts are needed to score the different bids". In a plant, add the operator or maintenance lead who will use the system.
Ask for a price per phase, each ending with a go/no-go: scoping, proof of concept, pilot, rollout, then each year of operation for at least three years, with person-days, hardware, licences and model updates on separate lines. Our guide to industrial AI project cost explains the cost lines and how Law 23-12 allows firm and conditional tranches.
Which rules apply depends on the buyer. Law 23-12 covers the State, local authorities and public-law establishments, among others (Article 9). Public economic enterprises are not bound by its award procedures but must adopt their own, based on free access, equal treatment and transparency (Articles 13 and 5). This is a summary, not legal advice; confirm with your procurement department and counsel.
What we have seen at KABAS AI
We are a vendor, so weigh this section accordingly.
Our published projects show two things to write into an RFP. At ENAD Shymeca, the Virtual Lab implementation began with a discovery phase alongside the R&D team, before configuration, migration of historical formulation data and training; scoping can be priced as its own phase. The case study also describes the first Open Innovation contract between an AI startup and a national Algerian company, a contract form worth discussing with your procurement team. At ADFC, decades of formulation knowledge were spread across formats and locations, exactly the kind of fact a data annex should state.
If hosting on your own infrastructure is required, as it can be with DALILE or Virtual Lab, say so in the scope, because it changes every bid. And if a vendor helps prepare the RFP, including us, decide beforehand whether it may also bid, and have someone independent check that the specification is not written around one product.
Our consulting team works on readiness assessment and use case identification, the groundwork for a problem statement and data annex. To discuss an RFP you are preparing, contact us.
Sources
- UK Office for Artificial Intelligence, Guidelines for AI procurement, 8 June 2020
- World Economic Forum, AI Procurement in a Box: AI Government Procurement Guidelines, June 2020
- Law 23-12 of 5 August 2023 on public procurement, Journal officiel no. 51 (FAOLEX copy)
- Gide, Protection of personal data: focus on Algerian regulations
- CISA, ASD's ACSC et al., Principles for the Secure Integration of Artificial Intelligence in Operational Technology, 3 December 2025
- OWASP, Top 10 for LLM Applications 2025
- NIST, AI Risk Management Framework
- ISO/IEC 42001:2023, Artificial intelligence management system (SIS catalogue entry)
- Mitchell et al., Model Cards for Model Reporting
- EU Community of Practice on Public Procurement of AI, Updated EU AI model contractual clauses, 5 March 2025
- Intelligent CIO, Gartner predicts 30% of generative AI projects will be abandoned after proof of concept by end of 2025
Frequently asked questions
What is the difference between an RFP and a cahier des charges?
RFP, or request for proposals, is the usual English term for the document that asks vendors for proposals. In Algerian public procurement, the tender documents are the cahiers des charges, which under Article 17 of Law 23-12 include general administrative clauses, common technical specifications and special specifications for each contract. The advice in this guide applies to either.
Should an AI RFP name a specific model or product?
No. Describe the problem, the constraints such as on-premise hosting or languages, and the performance to reach, and let bidders propose the technology. For buyers under Law 23-12, Article 16 requires technical specifications that are not oriented towards a given product or supplier. You can still set hard constraints, such as running without an internet connection.
Should we share real data with bidders?
Yes, a representative sample, under written conditions. The World Economic Forum guidelines suggest sharing only a sample when data is sensitive, limiting its use to the bid, and setting a deletion date with written confirmation. Give the same sample to every bidder at the same time, remove personal data, and check with counsel before sending anything abroad.
How do we compare a cloud offer with an on-premise offer?
Compare them on the same acceptance test and on total cost over at least three years, including hardware, licences, staff time and model updates. First check whether your data may leave the site at all. If it may not, the cloud offer fails a hard requirement, whatever its price.
Can a vendor that helped write the RFP also bid?
It depends on your procurement rules, so ask your procurement department and counsel. If it is allowed, make sure the specification is not oriented towards that vendor's product, give all bidders the same information, and have someone independent review the requirements.
Do private companies have to follow Law 23-12?
No. Law 23-12 applies mainly to the State, local authorities and public establishments. Public economic enterprises must adopt their own procedures based on its principles of free access, equal treatment and transparency. Private companies set their own rules, but published criteria and equal treatment of bidders still produce better, comparable bids.



