Winning the tender: how a 340-person government advisory firm out-bid firms 10× its size.
Orchestrary deployed nine production agents on OpenClaw — running fully on the firm's own infrastructure — and trained 38 consultants to extend them. Tender-response throughput rose 4.8×, win rate climbed from 22% to 38%, and the firm added €3.6M in additional won contracts in nine months.
The client is one of Central Europe's most respected public-sector advisory firms. Anonymized here, but recognizable to anyone in the regional consulting market: 340 people, three offices, 28 years of history, a portfolio that includes ministry-level digital strategy, EU structural-fund program design, and large-scale public procurement support. They are routinely shortlisted against the regional offices of the Big Four.
By 2024 they had a problem. Public-sector tendering had industrialized. EU procurement processes — particularly under the new procurement directive and the Horizon and Digital Europe instruments — required tender responses of 200–400 pages, structured to extremely specific evaluation criteria, with full evidence packs, methodology annexes, CV books, financial models, and compliance attestations. The firm's senior partners were spending 60% of their time on bid production. Their hit rate was respectable (22%) but their throughput was capped: they simply could not respond to enough opportunities to grow.
Meanwhile, three competitors had announced "AI-assisted bidding" services. The firm's leadership was rightly skeptical of the marketing, but they understood the implication: the firms that figured out agent-assisted tendering first would compound a margin and velocity advantage that the others would not catch up to.
They engaged Orchestrary on two non-negotiable conditions:
- Sovereignty. No client material, no tender draft, no internal know-how could ever leave the firm's own infrastructure. The agents had to run on the firm's own servers, in its own cloud tenant, against models the firm controlled. This ruled out anything built on hosted commercial AI agents.
- Independence. After the engagement, the firm had to be able to extend, modify, and operate the agents without us. No SaaS subscription. No vendor lock-in.
Both conditions pointed to the same toolchain: OpenClaw, the open-source, on-premise-deployable agent CLI we use for sovereignty-sensitive clients, running against a self-hosted Llama-class model with a routed fallback to a EU-region commercial endpoint for the highest-complexity tasks.
The constraint: why "ChatGPT for proposals" was never going to work
The firm had already tried the obvious thing. Two senior partners had piloted ChatGPT and Claude.ai in late 2023 for tender drafting. They abandoned the experiment within six weeks for three reasons:
- Confidentiality. Pasting client material into a hosted chat interface was a procurement-clause violation in roughly half of the firm's tenders.
- Quality. The output was generic. Public-sector tender writing is a specific genre — methodology sections must reference specific EU regulations and named programs, CV books must follow narrow formatting rules, financial models must align to specific cost-category schemas. A general-purpose chatbot produced text that sounded plausible and was uniformly wrong.
- Process integration. A consultant typing prompts into a browser tab is not a workflow. It does not version, it does not collaborate, it does not integrate with the firm's bid library, and it does not learn from the last 200 winning bids.
The lesson was clear. The firm did not need a chatbot. It needed agents that lived inside the firm's bid production system, knew the firm's win library, ran on the firm's own machines, and could be operated by consultants the same way an associate operates a junior team.
OpenClaw fit. We started in week three.
The portfolio: nine agents covering the full bid lifecycle
Wave 1 — Built by Orchestrary
TENDER-SCOUT · Opportunity identification and qualification
- Coverage: from ~80% to 99.4% within 4 hours of publication
- Analyst time: 50% of two FTEs → 6% of one FTE (~340 hours/month freed)
- Pursued opportunities: 87 → 162 (+86%) in the comparable 9-month window
BID-DRAFTER · Methodology and approach section drafting
- Time per methodology: 4–6 days → 6–8 hours of consultant review (85%+ reduction)
- First-draft quality: 78% of paragraphs accepted unchanged or with minor edits
- Consistency: 94% adherence to house style (vs. 60–70%)
CV-COMPOSER · CV book assembly and tailoring
- Time per CV book: 8–14 hours → 35 minutes (95%+ reduction)
- Expert utilization: 14 named experts under-deployed; their bid inclusion rate now 3.2× higher
- Compliance errors: dropped from ~6% of submissions to 0
COMPLIANCE-CHECKER · Pre-submission compliance and completeness audit
- Procedural disqualifications: 4.0% → 0.2% of submissions
- Estimated direct savings: €380K per year
- Pre-submission stress: dramatically reduced last-72-hour anxiety on submission week
REFERENCE-MATCHER · Past project reference selection and write-up
- Time per reference set: 4–8 hours → 25 minutes
- Memory: surfaces strong references from projects more than 15 years old that no current consultant remembered
Wave 2 — Built by the firm's consultants using OpenClaw (months 5–14)
After completing Academy tracks 1–3, four teams of consultants began building agents themselves. We provided weekly office hours but did not lead the work.
PRICING-MODELER · Q&A-RESPONDER · POST-MORTEM-LEARNER · REGULATORY-WATCHER
Four additional agents shipped between months 5 and 14, each authored entirely by Academy graduates. PRICING-MODELER cut tender financial modeling from 2 days to 90 minutes. Q&A-RESPONDER drafts answers to procurement clarification questions in 45 minutes. POST-MORTEM-LEARNER ingests every loss-feedback letter and produces quarterly pattern reports — already responsible for two strategy shifts visible in the firm's win rate. REGULATORY-WATCHER replaced an external regulatory monitoring subscription that had cost €34,000 per year.
The aggregate impact: bidding as a strategic advantage
| Metric | Year before | Year of (annualized at month 9) | Change |
|---|---|---|---|
| Tenders identified per quarter | 142 | 287 | +102% |
| Tenders pursued per quarter | 24 | 47 | +96% |
| Bid throughput per consultant per quarter | 0.6 | 2.9 | +383% |
| Win rate | 22% | 38% | +73% |
| Procedural disqualifications | 4.0% | 1.6% | −95% |
| Won contract value (12-mo rolling) | €41.2M | €43.4M | +€3.6M |
| Bid production cost per pursued opp. | €38,400 | €14,200 | −63% |
Senior partners' time on bid production: 60% → 22% (freed for client delivery and BD). Average bid team size: 6.4 people → 2.9 people. Bid library reuse rate: 34% → 68%.
Strategic outcome
The firm now competes credibly for tenders 2–3× larger than its historical sweet spot, because it can mount the response volume and quality these larger procurements demand. In the most recent reporting quarter, the firm was shortlisted in two opportunities where the only other shortlisted firm was a Big Four office.
What "deploy AND teach with OpenClaw" actually meant
1. Everything ran on the firm's own infrastructure
The agents, the model endpoints, the bid library indices, every artifact: all of it ran in the firm's own cloud tenant, behind the firm's own VPN, under the firm's own access controls. The firm's General Counsel was able to give an unqualified green light to using the system on every tender, including those subject to the strictest confidentiality clauses (defense, intelligence services, judicial reform).
2. The firm owns the entire stack
There is no Orchestrary subscription. There is no recurring license. The firm's CTO has root access to every component. He has, on three occasions, modified core OpenClaw configuration files himself.
3. The Academy turned consultants into agent operators
The Academy is now run internally. Cohort 4 (currently in progress) is being delivered by a Cohort 1 graduate.
The human dimension
"I have been doing public-sector tendering for 24 years. The shape of the work has changed twice in that time — once when the EU procurement directive came in, and again with this engagement. We are competing against firms ten times our size and winning. That was not possible in 2023."
Managing Partner
"I used to dread the four weeks before a major submission. Now I look at a tender notice and I know we can mount a serious response in five days. The agents do not write the bid — I still write the bid. They give me a draft I can react to instead of a blank page. The leverage is unbelievable."
Senior Bid Manager
"I was the person who killed the ChatGPT pilot. I have signed off on every aspect of the OpenClaw system. The difference is straightforward: this one runs on our hardware and our staff understand it from the inside."
General Counsel
Lessons for other government and public-sector advisory firms
When this model fits
- You operate in regulated procurement environments where data confidentiality is non-negotiable
- Your bid production is the bottleneck on growth, not your delivery capacity
- You have a deep historical bid and project archive that you currently cannot fully exploit
- You face competition from much larger firms who you can beat on quality but lose to on volume
- You want a permanent in-house capability, not a SaaS dependency
The deliverable
The contracted commitment was nine months. We exited on schedule. The firm's "Agent Engineering" function is now a permanent four-person internal team, all Academy graduates. Their roadmap includes seven additional agents over the next 18 months — none of which we will build.
We did the first wave. We taught the discipline. The firm owns the capability. That is the engagement.
Reading faster than the field: 14 researchers at CIIRC, 3.2× publications, 47 hypotheses
Want to be the seventh case study?
A 60-minute discovery call. No software pitch. We map your most painful workflow, scope a first agent, and tell you honestly whether this engagement model fits your organization.