United Nations Office for Project Services: Creating an AI playbook

The problem

The United Nations Office for Project Services (UNOPS) is the implementation arm of the UN system, delivering around 1,000 projects a year worldwide. These range from installing wells to building multiple hospitals within a single country. UNOPS works in some of the most challenging environments in the world, including conflict zones and hard-to-reach areas, and also manages large-scale public procurement and distributes grants on behalf of other organisations to local NGOs and implementing partners.

UNOPS operates under a complex funding model. Rather than receiving core funding, it charges a management fee as a percentage of each project's cost. This means the organisation competes for work and has a strong incentive to keep overheads low.

Despite managing projects worth billions of dollars collectively, UNOPS never has more than about 6,000 staff at any one time, and that number fluctuates as people join for specific projects and leave when those projects end. Staff wear many hats, need to be onboarded quickly, and face heavy administrative workloads across procurement, grant management, infrastructure oversight, and business development.

Many of these processes are labour-intensive. Evaluating grant proposals, for example, could involve reviewing submissions from 200 implementing partners, a process that would normally take months. Headquarters review of individual engagements requires checking against extensive compliance checklists covering IT security, physical security, financing, procurement, and health, safety, and environmental standards, a process that involves weeks of communication between headquarters and field teams. In infrastructure projects in remote or conflict-affected areas, project managers are sometimes unable to physically visit construction sites to check that health and safety standards are being met.

The solution

UNOPS developed AI Playbook, a web-based platform that allows staff across the organisation to create and use AI-powered templates to automate specific tasks and processes.

A playbook is a reusable AI workflow. It is similar to a structured template, checklist, or guided form, but with AI built into it. Instead of asking staff to start from a blank prompt, a playbook packages together the task instructions, the documents or data the AI should review, the organisational policies or criteria it should apply, and the format of the output it should produce.

In practical terms, a playbook answers four questions for the user:

1. What task is being done? For example, screening grant proposals, comparing vendor quotations, checking a budget, or reviewing construction site photos.
2. What information does the AI need? This might include uploaded proposals, quotations, photographs, budgets, project documents, or policy files.
3. What rules or standards should the AI use? The playbook can be preloaded with UNOPS policies, eligibility criteria, checklists, calculators, or health and safety guidelines.
4. What output should the user receive? The playbook might produce a gap analysis, a comparison table, a preliminary eligibility screen, a list of safety concerns, or a budget error report.

This means a staff member does not need to design a new AI prompt every time they want help with a recurring task. They can open the relevant playbook, add the required inputs, and receive a structured, AI-generated output ready for human review.

Staff access the AI Playbook through their browser. For simpler use cases, staff can create their own playbooks by defining the task, adding the relevant instructions and reference materials, and setting the desired output format. For more complex processes, such as grant evaluation or procurement, the product team builds, tests, and validates the product, then onboards the relevant team to use the playbook.

UNOPS built the platform on Google technologies, including Gemini. It also developed an internal contextual knowledge base, known as BOB, which is a retrieval-augmented generation system. Retrieval-augmented generation, or RAG, is an approach in which the AI draws its answers from a specific set of organisational documents and data rather than relying only on general information. This helps ensure that answers are grounded in UNOPS’s own policies, rules, and context.

Over approximately five to six months, AI Playbook evolved from automating individual tasks to automating more complete processes. In some cases, the output of one stage can feed automatically into the next stage, creating a connected workflow rather than a one-off AI interaction.

The AI Playbook supports a range of use cases across UNOPS.

Procurement
One of the first playbooks automated a process known as “shopping,” which refers to the procurement of lower-value items. Quotations from vendors can arrive in many formats, including scanned documents, system-generated quotations, photographs of products, and handwritten notes. The playbook processes these materials, compares them, and produces an evaluation. This reduces manual effort in a high-volume process that spans the organisation.

Grant management
UNOPS receives proposals from hundreds of implementing partners. Checking whether each proposal meets mandatory eligibility requirements, such as whether an audited financial statement has been included, previously required expert reviewers to work through every submission individually. The playbook now completes preliminary and eligibility screening in minutes, giving evaluators a head start so they can focus on the substantive technical assessment.

Headquarters oversight
For headquarters review, the platform checks project documentation against organisational policies and standards. It can identify gaps in areas such as IT security, physical security, financing, procurement, and health, safety, social, and environmental requirements. This provides headquarters and field teams with an initial gap analysis that would previously have required weeks of back-and-forth communication.

Health and Safety monitoring
At construction sites in conflict zones or difficult-to-reach areas, contractors upload photographs at regular intervals. The AI checks these images against health and safety guidelines and flags potential issues, such as missing safety equipment, inadequate fencing, risky excavation, or absent signage. This does not replace site visits or the project manager's responsibility, but it provides an additional form of oversight when physical visits are constrained.

Budget review
Staff can enter project budgets into a playbook preloaded with policies, guidelines, and calculators. The playbook identifies errors such as incorrect conversion rates or discrepancies between staffing plans and budget lines. This gives staff faster feedback before budgets move further through the review process.

Across all use cases, the AI acts as a smart and context-aware assistant, not a decision-maker. It does not make final decisions or take autonomous action. Human staff review the outputs and remain responsible for judgement, approval, and accountability.

The impact

1. An estimated 14,000 staff hours saved per year from a single process alone
The shopping procurement playbook alone is estimated to save around 14,000 human hours per year, based on the volume of shopping activity across the organisation. This is for a high-volume but relatively uncomplicated process, suggesting potential for further savings in more complex use cases.

2. Grant proposal screening reduced from months to minutes
Preliminary and eligibility screening of grant proposals, which is largely objective, previously required expert reviewers to work through each submission individually over months, can now be completed in minutes. Technical evaluation still requires human judgement, but reviewers begin the process significantly further ahead.

3. Headquarters review processes shortened from weeks to minutes
Engagement reviews and compliance checks against organisational standards and policies, which previously took weeks of communication between headquarters and field teams, can now produce an initial gap analysis in minutes.

4. Remote construction site monitoring made possible
In conflict-affected or hard-to-reach areas where project managers cannot visit construction sites, AI-powered analysis of photographs helps identify health and safety risks that would otherwise go undetected between visits.

Lessons learned

• Start small and build momentum. The team deliberately chose not to begin with the biggest or most impactful use case. Instead, they selected smaller, well-defined tasks where they had stakeholder backing and engaged teams who would take ownership. Early successes created interest across the organisation, with other teams requesting their own playbooks.
• Design AI as an assistant. Across all use cases, the platform is positioned as a tool that supports human decision-making rather than replacing it. This was important for maintaining trust and ensuring that the AI operates within the organisation's governance frameworks.
• Lean teams can deliver if the approach is right. The AI Playbook was developed with an extremely lean setup compared to larger transformation programmes within the organisation. The combination of readily available AI capabilities and a focused product approach enabled rapid progress without large, dedicated budgets or teams.
• Onboarding and familiarity remain a challenge. While simpler playbooks can be created by staff independently, more complex ones require hands-on support, training materials, and video guides. Building organisational familiarity with the platform is an ongoing effort.