AI in contract work: a comparison using SSA-Lille Sky
Frode Stenstrøm and Adrian Sjøholt had Google Gemini, Microsoft Copilot and Claude compete at drafting a contract.
Frode Stenstrøm and Adrian Sjøholt had Google Gemini, Microsoft Copilot and Claude compete at drafting a contract.
Frode Stenstrøm and Adrian Sjøholt from ProsessPilotene tested Google Gemini 2.5, Microsoft Copilot and Anthropic's Claude Sonnet against each other in a live competition: filling in the appendices of the Norwegian public-sector standard agreement SSA-Lille Sky for a hypothetical cloud-service delivery. Both used the free versions where possible, gave the models the same initial prompt, and let the models ask their own follow-up questions before filling in the appendices.
Gemini and Claude both managed to produce full draft agreements of around 60 pages, with precise contract language, correctly structured appendices and, in places, independently filled-in details such as a responsibility matrix and cross-references to the main agreement. The free version of Copilot, however, did not finish: it turned out to have a context limit of only 4,000 tokens, against Gemini's one million, and so could not process the full contract material.
Along the way, Frode showed how Gemini's canvas feature lets you edit individual paragraphs without rewriting the whole document, and how the model could look up current links to data processing agreements and legal text on wage and working conditions. Claude Sonnet gave a clean presentation and good text quality, but started losing track of earlier edits once the document grew too long for its context, so parts had to be split off and updated separately.
The 2025 conclusion was that Gemini and Claude are close to equal in quality, but that free Gemini is a natural first choice since Claude requires a paid subscription. A contract draft that normally takes 40-60 hours could be produced in two to four hours, though the presenters stressed the result still needs to be reviewed by whoever is responsible for the agreement.
Gründer og teknologistrateg
Som gründer og teknologistrateg i ProsessPilotene jobber jeg med å hjelpe virksomheter forstå og bruke teknologi som en strategisk kraft for bedre arbeidsprosesser, kundeopplevelser og beslutninger. Rollen kombinerer forretningsutvikling, teknologirådgivning, virksomhetsarkitektur og formidling. Jeg arbeider særlig i skjæringspunktet mellom AI, CRM, digitale plattformer, data, analyse og fremtidens arbeidsliv. En sentral del av rollen er å være brobygger mellom ledelse, fagmiljøer og teknologi. Det handler om å gjøre komplekse teknologiske muligheter forståelige, vurdere hva som faktisk skaper verdi, og bidra til at digitale løsninger blir en naturlig del av hvordan virksomheter jobber. Jeg bruker også rollen til å formidle faktabaserte perspektiver på hvordan AI og teknologi påvirker ledelse, organisasjoner, ansatte og kunder.
CRM-Konsulent
Ask a question or share what helped you.
Frode Stenstrøm and Adrian Sjøholt from ProsessPilotene tested Google Gemini 2.5, Microsoft Copilot and Anthropic's Claude Sonnet against each other in a live competition: filling in the appendices of the Norwegian public-sector standard agreement SSA-Lille Sky for a hypothetical cloud-service delivery. Both used the free versions where possible, gave the models the same initial prompt, and let the models ask their own follow-up questions before filling in the appendices.
Gemini and Claude both managed to produce full draft agreements of around 60 pages, with precise contract language, correctly structured appendices and, in places, independently filled-in details such as a responsibility matrix and cross-references to the main agreement. The free version of Copilot, however, did not finish: it turned out to have a context limit of only 4,000 tokens, against Gemini's one million, and so could not process the full contract material.
Along the way, Frode showed how Gemini's canvas feature lets you edit individual paragraphs without rewriting the whole document, and how the model could look up current links to data processing agreements and legal text on wage and working conditions. Claude Sonnet gave a clean presentation and good text quality, but started losing track of earlier edits once the document grew too long for its context, so parts had to be split off and updated separately.
The 2025 conclusion was that Gemini and Claude are close to equal in quality, but that free Gemini is a natural first choice since Claude requires a paid subscription. A contract draft that normally takes 40-60 hours could be produced in two to four hours, though the presenters stressed the result still needs to be reviewed by whoever is responsible for the agreement.
Build an account plan the whole sales team can use
CoursePrioritise account work using evidenced needs, relevant relationships and valuable next actions.
Learning path: customer success, value and renewal
CourseGrow existing customer relationships through evidenced value, early risk response and renewals grounded in customer needs.
Make Copilot a useful part of customer work
CourseDesign and test a role-based AI workflow with clear sources, checks and accountability.
Use Copilot in Excel for analysis you can verify
CourseAsk precise data questions, check calculations and turn findings into decisions with visible qualifications.
Jets Group's experience with AI in sales and order work
October 21, 2025 • Webinar — Trond Jarnes of Jets Vacuum told Stine Sletvold and Frode Stenstrøm how AI in Dynamics 365 Sales now automates quote lines.
AI agents: use cases and implementation choices
May 16, 2025 • Webinar — Adrian Sjøholt and Frode Stenstrøm of ProsessPilotene covered trends, models and use cases for AI agents in business.
June 19, 2025 • Webinar — Frode Stenstrøm and Adrian Sjøholt tested Google, OpenAI and Anthropic AI models on real project-planning tasks, from requirements to a presentation.
April 2, 2025 • Webinar — Ane Tronsaune showed Frode Stenstrøm how to automate a Dynamics 365 discount approval with Power Automate, Teams and AI Builder.
Sales | Customer story
Kartverket got a new, tailor-made customer portal built on modern technology – a foundation that makes it easy to add new services, integrations and improvements as users' needs evolve.
Sales | Customer story
In six weeks, we built the AI agent “Oddie”. It identifies new business opportunities around the world for Odfjell Technology.
Right from the start of the project, great emphasis was placed on training super users and closely involving the organization. This has contributed to a solid rollout and strong buy-in across the sales and marketing teams.
Automation | Article
The data agent gained a larger role in analytics
OpenAI’s Data agent could investigate company data and build shareable interactive dashboards. Access controls and source trails needed to follow analysis and action through the workflow.
Customer Service | Article
Service Agent gained tools for staffing decisions
WEM MCP tools enabled time-off requests, shift swaps and shift bids in Service Agent. Administrators could enable each tool for a role, team or pilot group.
Automation | Article
Microsoft moved to a continuously updated roadmap
Microsoft replaced twice-yearly release waves with continuous roadmap publication. Customer processes needed to track status from development through rollout to launch.
Information security | Article
Norway to reconsult on its AI law after EU rule changes
On 4 August 2026, the Norwegian government announced that its proposed AI Act would return to consultation after changes in the EU simplification package.
Sales and Service Agent became generally available in Microsoft 365 Copilot. Dynamics described MCP actions across sales, service, field work and projects.
The EDPB published new guidelines on web scraping for generative AI and adopted final blockchain guidelines.
Share a question or reflection.
Be the first to contribute.
