Podcast Summary: "Building Tendos AI: How an Agent Swarm Turns Construction Emails into Quotes"

Podcast: Just Now Possible
Host: Teresa Torres
Guests: Daniel (CPO, Tendos), Matthias (CTO, Tendos)
Date: January 15, 2026

Episode Overview

This episode dives deep into the story behind Tendos AI, an agent-swarmed platform transforming how manufacturers in the construction industry manage and respond to incoming requests—particularly by turning complex, unstructured construction emails and documents into accurate quotes. Host Teresa Torres explores with Daniel and Matthias how the Tendos team identified pressing industry inefficiencies, validated the AI solution, approached prototyping, expanded their product scope, and built a robust, multi-agent system to streamline construction workflows. The discussion covers early bets on AI, the unique challenges of parsing multi-format documents, technical architecture, iterative evaluation, and user adoption subtleties within a traditionally difficult-to-digitize domain.

Key Discussion Points & Insights

1. Understanding Inefficiency in Construction Workflows

[00:32-04:12]

• Daniel: Tendos sits as a "system of action" above legacy ERP and CRM systems, targeting back-office workflows for manufacturers, where inefficiency and high volume of manual email traffic is rampant.

• Tendering chain complexity: Construction projects involve multiple parties (owners, architects, planners, manufacturers, wholesalers) and are not 1-to-1 relationships but vast, branching networks. Manual alignments and slow communication create enormous delays.

  • "You end up in like a huge tree of different parties involved and you have all these manual jobs...happening on the way." (Daniel, 01:09)

• Teresa: Shares a personal analogy about managing a house renovation, illustrating the fragmented, multi-party coordination problem—even at small household project scale.

2. The Pain & Potential of Offers and Bids

[04:12-07:24]

• A robust, detailed offer leads to customer trust and efficiency.

  • "We actually didn't go with them because we're like, we're terrified that you're just going to come back and charge us more money...I can definitely see how software can play a role." (Teresa, 04:12)

• Tendos AI's focus: Automate the prioritization, extraction, and structuring of key data from incoming emails and complex attachments (PDFs, Excels).

  • Extraction includes everything from customer intent and urgency to technical product specs.

• The primary workflow starts with incoming (sometimes unstructured) requests, categorizes and prioritizes them, extracts relevant product and partner information, and pre-populates draft offers for human review.

3. Deep Dive into How AI Powers the Product

[07:39-11:16]

• Matthias: The AI must:

  • Understand ambiguous human requests (which range from precise specs to vague intentions).
  • Map requests to the manufacturer’s actual product portfolio, checking not only availability but also fit and feasibility (e.g., if the customer wants copper pipes, and only plastic are available).
  • Document extraction via LLMs has made it easier to process previously unstructured, extensive attachments—"cutting them into pieces," extracting relevant positions from hundred-page documents.
  • "Most of the critical information...is semantics. It's very hard for computers to understand before the age of LLMs." (Matthias, 07:39)

• The system can automate simple offers but keeps a "human in the loop" for complex and high-stakes cases.

4. Product Origins: Prototyping and Early Customer Validation

[13:19-18:23]

• Tendos was founded out of direct personal experience with construction industry inefficiencies.

• The leap to AI was enabled by a believed "moment of change"—AI unlocking enterprise willingness to reimagine established workflows.

  • Early efforts included "very naive" prototypes: simply feeding emails and documents to LLMs (e.g., ChatGPT) to see if output matched human intent.
  • "We quickly found out that we actually can do that...that was the moment where we said, okay, we see the problem." (Matthias, 14:41)

• Initial user response was so positive that users anthropomorphized the system (gave it a name, "chassis") and quickly began requesting the AI system replace existing modules in legacy software (like SAP CPQ).

5. Execution: Narrow Start, Demand-Led Expansion

[19:12-32:08]

• First "bite of the apple": The team started with a single, narrow product line (specialty radiators) with a design partner company, isolating complexity and validating the solution in a contained environment.

• This hyper-focus built product quality and confidence before gradual expansion:

  • From simple, single-product PDF requests.
  • To larger, multi-product documents.
  • Eventually to parsing out relevant line items from massive, building-spanning documentation (up to nearly 2000 pages).
  • "Our users started to experiment and the documents grew larger and larger, to the point that...now we need to invest more into entity extraction." (Matthias, 24:04)

• Sitting alongside end users, observing real-world usage, surfaced countless adjacent inefficiencies and expanded the scope of automation.

6. Scaling to a Robust Multi-Agent System

[38:21-65:21]

• Current product: Handles the full workflow from email intake (support, offer, order), auto-categorizing, prioritizing, tying to CRM/project, extracting request content, and generating draft offers/support responses for human approval.

• Technical details:

  • Static early pipeline steps: Email parsing, context lookup, entity extraction.
  • Agentic and dynamic execution: Based on findings, a "planning" module dynamically orchestrates LLM-powered agents for contextual reasoning, product matching, and quality assurance.
  • Review/critic agents: Specialized agents evaluate proposed matches, semi-autonomously "reviewing" work akin to code review, prompting corrections and reruns as needed.
  • The system "knows when it doesn't know" and can pass uncertain matches back to a human to prevent costly errors.
    • "If you're looking at a product portfolio and you want to match a product to a description, then there's not a lot of value in saying, hey, how confident am I?...We'd rather say we don't know yet..." (Matthias, 39:27)

• Eval & feedback: Heavy investment into step-level evaluation and tracing—each agent decision is checked, and both user feedback and production data are used to refine performance.

  • Proprietary tooling was built for traceability, error analysis, and querying across evaluation results.

7. Approach to Product Expansion & Culture

[34:29-36:28]

• Tendos deliberately avoids overextending, focusing tightly on segments where their agentic automation delivers highest value (e.g., avoiding road-building projects, which involve entirely different workflows/data).

  • "Focus is key...solving the last 20% is really hard." (Daniel, 34:58)

• They eschew scaling support teams, instead prioritizing engineering and AI-driven automation of support and new workflows.

8. The Future: More Agents, Broader Workflows, Deeper Customer Integration

[63:15-65:21]

• Rapid expansion is customer-demand driven, but always tempered by strategic focus and careful, stepwise product evolution.
• Next steps include: growing engineering team, more deeply integrated user experiences, more flexible and adaptable pipelines, and continued modularity and quality in the agent swarm.
  • "More agents, more problems to solve...growing engineering team because more and more customer requests are coming in..." (Matthias, 63:16)
  • Exploring more dynamic customization and interaction patterns, and higher-order automation without sacrificing user trust or product quality.

Notable Quotes & Memorable Moments

• On the scale and challenge of construction tendering:
  • "It's not a one-to-one relationship but a one-to-end problem. You end up in like a huge tree of different parties involved." (Daniel, 01:09)
• On why AI is the unlock:
  • "AI is the first point in time where now again enterprises are open to rethink their whole tool lens." (Daniel, 13:19)
• The charm of user adoption:
  • "...after two days using our platform, the users were doing an internal workshop to give a name [to] our AI system..." (Daniel, 16:39)
• On agentic architecture:
  • "Whatever that means for anyone, it's an agentic architecture..." (Matthias, 47:17)
• Confidence and uncertainty in AI product matching:
  • "...we'd rather say we don't know yet and show a choice than to say, ah, we're somewhat certain and this is what we came up with." (Matthias, 39:27)
• On product and market focus:
  • "Focus is key...it's easy to get to 70-80%, but...solving the last 20% is really hard." (Daniel, 34:58)
• On evaluating and tracing agent chains:
  • "Figuring out how to evaluate each agent in this application was quite tricky...But we see it as vital...it makes debugging so much easier..." (Matthias, 59:37)
• On future direction:
  • "More agents, more problems to solve...We are trying to expand further..." (Matthias, 63:16)

Timestamps for Key Segments

• Intro to Tendos & Industry Challenge: 00:27–02:35
• Teresa’s Construction Analogy: 02:35–04:12
• Manual Offer Creation Problems: 04:12–07:24
• How AI Powers Document/Email Understanding: 07:39–11:16
• Product Origins & Finding Early PMF: 13:19–18:23
• Narrow Initial Prototype & Product Evolution: 19:12–32:08
• Scaling Complexity: From PDFs to Swarm Agents: 32:08–38:21
• Architecture Deep Dive (Agentic Workflows): 38:53–54:39
• Evaluation, Tracing, and Agent Review: 59:18–62:45
• Future Expansion & Closing Thoughts: 63:15–65:21

Summary Takeaways

Tendos AI’s story stands out as an archetypal example of vertical AI product-building: starting with deep, founder-driven domain expertise; narrowly scoping the first use case; and using both technological advancements and close user partnership to iteratively expand. At every turn, the Tendos team balances customer pull with a rigorous insistence on product focus and technical depth, building a robust, multi-agent system capable of handling real-world, high-stakes construction workflows.

If you're building AI-driven automation for manual, high-variance domains, or seeking lessons on human-in-the-loop, agentic architectures, or evaluating AI at scale, this episode is a treasure trove.